fix: add missing session parameters to controller method calls

- Add db_session parameter to all RoomController.add() and update() calls in test_room_ics_api.py
- Fix TranscriptController.upsert_topic() calls to include session parameter in conftest.py fixture
- Fix TranscriptController.upsert_participant() and delete_participant() calls to include session parameter in API views
- Remove invalid setup_database fixture references, use pytest-async-sqlalchemy's database fixture instead
- Update CalendarEventController.upsert() calls to include session parameter

These changes ensure all controller methods receive the required session parameter
as part of the SQLAlchemy 2.0 migration pattern where sessions are explicitly managed.

.github/workflows/db_migrations.yml

@@ -2,6 +2,8 @@ name: Test Database Migrations
 
 on:
   push:
+    branches:
+      - main
     paths:
       - "server/migrations/**"
       - "server/reflector/db/**"
@@ -17,10 +19,43 @@ on:
 jobs:
   test-migrations:
     runs-on: ubuntu-latest
+    concurrency:
+      group: db-ubuntu-latest-${{ github.ref }}
+      cancel-in-progress: true
+    services:
+      postgres:
+        image: postgres:17
+        env:
+          POSTGRES_USER: reflector
+          POSTGRES_PASSWORD: reflector
+          POSTGRES_DB: reflector
+        ports:
+          - 5432:5432
+        options: >-
+          --health-cmd pg_isready -h 127.0.0.1 -p 5432
+          --health-interval 10s
+          --health-timeout 5s
+          --health-retries 5
+
+    env:
+      DATABASE_URL: postgresql://reflector:reflector@localhost:5432/reflector
 
     steps:
       - uses: actions/checkout@v4
 
+      - name: Install PostgreSQL client
+        run: sudo apt-get update && sudo apt-get install -y postgresql-client | cat
+
+      - name: Wait for Postgres
+        run: |
+          for i in {1..30}; do
+            if pg_isready -h localhost -p 5432; then
+              echo "Postgres is ready"
+              break
+            fi
+            echo "Waiting for Postgres... ($i)" && sleep 1
+          done
+
       - name: Install uv
         uses: astral-sh/setup-uv@v3
         with:

.github/workflows/deploy.yml

@@ -8,18 +8,30 @@ env:
   ECR_REPOSITORY: reflector
 
 jobs:
-  deploy:
-    runs-on: ubuntu-latest
+  build:
+    strategy:
+      matrix:
+        include:
+          - platform: linux/amd64
+            runner: linux-amd64
+            arch: amd64
+          - platform: linux/arm64
+            runner: linux-arm64
+            arch: arm64
+
+    runs-on: ${{ matrix.runner }}
 
     permissions:
-      deployments: write
       contents: read
 
+    outputs:
+      registry: ${{ steps.login-ecr.outputs.registry }}
+
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
 
       - name: Configure AWS credentials
-        uses: aws-actions/configure-aws-credentials@0e613a0980cbf65ed5b322eb7a1e075d28913a83
+        uses: aws-actions/configure-aws-credentials@v4
         with:
           aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
           aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
@@ -27,21 +39,52 @@ jobs:
 
       - name: Login to Amazon ECR
         id: login-ecr
-        uses: aws-actions/amazon-ecr-login@62f4f872db3836360b72999f4b87f1ff13310f3a
-
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
+        uses: aws-actions/amazon-ecr-login@v2
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
 
-      - name: Build and push
-        id: docker_build
-        uses: docker/build-push-action@v4
+      - name: Build and push ${{ matrix.arch }}
+        uses: docker/build-push-action@v5
         with:
           context: server
-          platforms: linux/amd64,linux/arm64
+          platforms: ${{ matrix.platform }}
           push: true
-          tags: ${{ steps.login-ecr.outputs.registry }}/${{ env.ECR_REPOSITORY }}:latest
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          tags: ${{ steps.login-ecr.outputs.registry }}/${{ env.ECR_REPOSITORY }}:latest-${{ matrix.arch }}
+          cache-from: type=gha,scope=${{ matrix.arch }}
+          cache-to: type=gha,mode=max,scope=${{ matrix.arch }}
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}
+          provenance: false
+
+  create-manifest:
+    runs-on: ubuntu-latest
+    needs: [build]
+
+    permissions:
+      deployments: write
+      contents: read
+
+    steps:
+      - name: Configure AWS credentials
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          aws-access-key-id: ${{ secrets.AWS_ACCESS_KEY_ID }}
+          aws-secret-access-key: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+          aws-region: ${{ env.AWS_REGION }}
+
+      - name: Login to Amazon ECR
+        uses: aws-actions/amazon-ecr-login@v2
+
+      - name: Create and push multi-arch manifest
+        run: |
+          # Get the registry URL (since we can't easily access job outputs in matrix)
+          ECR_REGISTRY=$(aws ecr describe-registry --query 'registryId' --output text).dkr.ecr.${{ env.AWS_REGION }}.amazonaws.com
+
+          docker manifest create \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest-amd64 \
+            $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest-arm64
+
+          docker manifest push $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest
+
+          echo "✅ Multi-arch manifest pushed: $ECR_REGISTRY/${{ env.ECR_REPOSITORY }}:latest"

.github/workflows/pre-commit.yml

@@ -0,0 +1,24 @@
+name: pre-commit
+
+on:
+  pull_request:
+  push:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: actions/setup-python@v5
+      - uses: pnpm/action-setup@v4
+        with:
+          version: 10
+      - uses: actions/setup-node@v4
+        with:
+          node-version: 22
+          cache: "pnpm"
+          cache-dependency-path: "www/pnpm-lock.yaml"
+      - name: Install dependencies
+        run: cd www && pnpm install --frozen-lockfile
+      - uses: pre-commit/action@v3.0.1

.github/workflows/test_next_server.yml

@@ -0,0 +1,45 @@
+name: Test Next Server
+
+on:
+  pull_request:
+    paths:
+      - "www/**"
+  push:
+    branches:
+      - main
+    paths:
+      - "www/**"
+
+jobs:
+  test-next-server:
+    runs-on: ubuntu-latest
+
+    defaults:
+      run:
+        working-directory: ./www
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Setup Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+
+    - name: Install pnpm
+      uses: pnpm/action-setup@v4
+      with:
+        version: 8
+
+    - name: Setup Node.js cache
+      uses: actions/setup-node@v4
+      with:
+        node-version: '20'
+        cache: 'pnpm'
+        cache-dependency-path: './www/pnpm-lock.yaml'
+
+    - name: Install dependencies
+      run: pnpm install
+
+    - name: Run tests
+      run: pnpm test

.github/workflows/test_server.yml

@@ -5,12 +5,17 @@ on:
     paths:
       - "server/**"
   push:
+    branches:
+      - main
     paths:
       - "server/**"
 
 jobs:
   pytest:
     runs-on: ubuntu-latest
+    concurrency:
+      group: pytest-${{ github.ref }}
+      cancel-in-progress: true
     services:
       redis:
         image: redis:6
@@ -19,29 +24,47 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Install uv
-        uses: astral-sh/setup-uv@v3
+        uses: astral-sh/setup-uv@v6
         with:
           enable-cache: true
           working-directory: server
-
       - name: Tests
         run: |
           cd server
           uv run -m pytest -v tests
 
-  docker:
-    runs-on: ubuntu-latest
+  docker-amd64:
+    runs-on: linux-amd64
+    concurrency:
+      group: docker-amd64-${{ github.ref }}
+      cancel-in-progress: true
     steps:
       - uses: actions/checkout@v4
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v2
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
-      - name: Build and push
-        id: docker_build
-        uses: docker/build-push-action@v4
+        uses: docker/setup-buildx-action@v3
+      - name: Build AMD64
+        uses: docker/build-push-action@v6
         with:
           context: server
-          platforms: linux/amd64,linux/arm64
-          cache-from: type=gha
-          cache-to: type=gha,mode=max
+          platforms: linux/amd64
+          cache-from: type=gha,scope=amd64
+          cache-to: type=gha,mode=max,scope=amd64
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}
+
+  docker-arm64:
+    runs-on: linux-arm64
+    concurrency:
+      group: docker-arm64-${{ github.ref }}
+      cancel-in-progress: true
+    steps:
+      - uses: actions/checkout@v4
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Build ARM64
+        uses: docker/build-push-action@v6
+        with:
+          context: server
+          platforms: linux/arm64
+          cache-from: type=gha,scope=arm64
+          cache-to: type=gha,mode=max,scope=arm64
+          github-token: ${{ secrets.GHA_CACHE_TOKEN }}

.gitignore

@@ -13,3 +13,8 @@ restart-dev.sh
 data/
 www/REFACTOR.md
 www/reload-frontend
+server/test.sqlite
+CLAUDE.local.md
+www/.env.development
+www/.env.production
+.playwright-mcp

.gitleaksignore

@@ -0,0 +1 @@
+b9d891d3424f371642cb032ecfd0e2564470a72c:server/tests/test_transcripts_recording_deletion.py:generic-api-key:15

.pre-commit-config.yaml

@@ -3,10 +3,10 @@
 repos:
   - repo: local
     hooks:
-      - id: yarn-format
-        name: run yarn format
+      - id: format
+        name: run format
         language: system
-        entry: bash -c 'cd www && yarn format'
+        entry: bash -c 'cd www && pnpm format'
         pass_filenames: false
         files: ^www/
 
@@ -23,8 +23,12 @@ repos:
       - id: ruff
         args:
           - --fix
-          - --select
-          - I,F401
+          # Uses select rules from server/pyproject.toml
         files: ^server/
       - id: ruff-format
         files: ^server/
+
+  - repo: https://github.com/gitleaks/gitleaks
+    rev: v8.28.0
+    hooks:
+      - id: gitleaks

CHANGELOG.md

@@ -1,5 +1,185 @@
 # Changelog
 
+## [0.13.1](https://github.com/Monadical-SAS/reflector/compare/v0.13.0...v0.13.1) (2025-09-22)
+
+
+### Bug Fixes
+
+* TypeError on not all arguments converted during string formatting in logger ([#667](https://github.com/Monadical-SAS/reflector/issues/667)) ([565a629](https://github.com/Monadical-SAS/reflector/commit/565a62900f5a02fc946b68f9269a42190ed70ab6))
+
+## [0.13.0](https://github.com/Monadical-SAS/reflector/compare/v0.12.1...v0.13.0) (2025-09-19)
+
+
+### Features
+
+* room form edit with enter ([#662](https://github.com/Monadical-SAS/reflector/issues/662)) ([47716f6](https://github.com/Monadical-SAS/reflector/commit/47716f6e5ddee952609d2fa0ffabdfa865286796))
+
+
+### Bug Fixes
+
+* invalid cleanup call ([#660](https://github.com/Monadical-SAS/reflector/issues/660)) ([0abcebf](https://github.com/Monadical-SAS/reflector/commit/0abcebfc9491f87f605f21faa3e53996fafedd9a))
+
+## [0.12.1](https://github.com/Monadical-SAS/reflector/compare/v0.12.0...v0.12.1) (2025-09-17)
+
+
+### Bug Fixes
+
+* production blocked because having existing meeting with room_id null ([#657](https://github.com/Monadical-SAS/reflector/issues/657)) ([870e860](https://github.com/Monadical-SAS/reflector/commit/870e8605171a27155a9cbee215eeccb9a8d6c0a2))
+
+## [0.12.0](https://github.com/Monadical-SAS/reflector/compare/v0.11.0...v0.12.0) (2025-09-17)
+
+
+### Features
+
+* calendar integration ([#608](https://github.com/Monadical-SAS/reflector/issues/608)) ([6f680b5](https://github.com/Monadical-SAS/reflector/commit/6f680b57954c688882c4ed49f40f161c52a00a24))
+* self-hosted gpu api ([#636](https://github.com/Monadical-SAS/reflector/issues/636)) ([ab859d6](https://github.com/Monadical-SAS/reflector/commit/ab859d65a6bded904133a163a081a651b3938d42))
+
+
+### Bug Fixes
+
+* ignore player hotkeys for text inputs ([#646](https://github.com/Monadical-SAS/reflector/issues/646)) ([fa049e8](https://github.com/Monadical-SAS/reflector/commit/fa049e8d068190ce7ea015fd9fcccb8543f54a3f))
+
+## [0.11.0](https://github.com/Monadical-SAS/reflector/compare/v0.10.0...v0.11.0) (2025-09-16)
+
+
+### Features
+
+* remove profanity filter that was there for conference ([#652](https://github.com/Monadical-SAS/reflector/issues/652)) ([b42f7cf](https://github.com/Monadical-SAS/reflector/commit/b42f7cfc606783afcee792590efcc78b507468ab))
+
+
+### Bug Fixes
+
+* zulip and consent handler on the file pipeline ([#645](https://github.com/Monadical-SAS/reflector/issues/645)) ([5f143fe](https://github.com/Monadical-SAS/reflector/commit/5f143fe3640875dcb56c26694254a93189281d17))
+* zulip stream and topic selection in share dialog ([#644](https://github.com/Monadical-SAS/reflector/issues/644)) ([c546e69](https://github.com/Monadical-SAS/reflector/commit/c546e69739e68bb74fbc877eb62609928e5b8de6))
+
+## [0.10.0](https://github.com/Monadical-SAS/reflector/compare/v0.9.0...v0.10.0) (2025-09-11)
+
+
+### Features
+
+* replace nextjs-config with environment variables ([#632](https://github.com/Monadical-SAS/reflector/issues/632)) ([369ecdf](https://github.com/Monadical-SAS/reflector/commit/369ecdff13f3862d926a9c0b87df52c9d94c4dde))
+
+
+### Bug Fixes
+
+* anonymous users transcript permissions ([#621](https://github.com/Monadical-SAS/reflector/issues/621)) ([f81fe99](https://github.com/Monadical-SAS/reflector/commit/f81fe9948a9237b3e0001b2d8ca84f54d76878f9))
+* auth post ([#624](https://github.com/Monadical-SAS/reflector/issues/624)) ([cde99ca](https://github.com/Monadical-SAS/reflector/commit/cde99ca2716f84ba26798f289047732f0448742e))
+* auth post ([#626](https://github.com/Monadical-SAS/reflector/issues/626)) ([3b85ff3](https://github.com/Monadical-SAS/reflector/commit/3b85ff3bdf4fb053b103070646811bc990c0e70a))
+* auth post ([#627](https://github.com/Monadical-SAS/reflector/issues/627)) ([962038e](https://github.com/Monadical-SAS/reflector/commit/962038ee3f2a555dc3c03856be0e4409456e0996))
+* missing follow_redirects=True on modal endpoint ([#630](https://github.com/Monadical-SAS/reflector/issues/630)) ([fc363bd](https://github.com/Monadical-SAS/reflector/commit/fc363bd49b17b075e64f9186e5e0185abc325ea7))
+* sync backend and frontend token refresh logic ([#614](https://github.com/Monadical-SAS/reflector/issues/614)) ([5a5b323](https://github.com/Monadical-SAS/reflector/commit/5a5b3233820df9536da75e87ce6184a983d4713a))
+
+## [0.9.0](https://github.com/Monadical-SAS/reflector/compare/v0.8.2...v0.9.0) (2025-09-06)
+
+
+### Features
+
+* frontend openapi react query ([#606](https://github.com/Monadical-SAS/reflector/issues/606)) ([c4d2825](https://github.com/Monadical-SAS/reflector/commit/c4d2825c81f81ad8835629fbf6ea8c7383f8c31b))
+
+
+### Bug Fixes
+
+* align whisper transcriber api with parakeet ([#602](https://github.com/Monadical-SAS/reflector/issues/602)) ([0663700](https://github.com/Monadical-SAS/reflector/commit/0663700a615a4af69a03c96c410f049e23ec9443))
+* kv use tls explicit ([#610](https://github.com/Monadical-SAS/reflector/issues/610)) ([08d88ec](https://github.com/Monadical-SAS/reflector/commit/08d88ec349f38b0d13e0fa4cb73486c8dfd31836))
+* source kind for file processing ([#601](https://github.com/Monadical-SAS/reflector/issues/601)) ([dc82f8b](https://github.com/Monadical-SAS/reflector/commit/dc82f8bb3bdf3ab3d4088e592a30fd63907319e1))
+* token refresh locking ([#613](https://github.com/Monadical-SAS/reflector/issues/613)) ([7f5a4c9](https://github.com/Monadical-SAS/reflector/commit/7f5a4c9ddc7fd098860c8bdda2ca3b57f63ded2f))
+
+## [0.8.2](https://github.com/Monadical-SAS/reflector/compare/v0.8.1...v0.8.2) (2025-08-29)
+
+
+### Bug Fixes
+
+* search-logspam ([#593](https://github.com/Monadical-SAS/reflector/issues/593)) ([695d1a9](https://github.com/Monadical-SAS/reflector/commit/695d1a957d4cd862753049f9beed88836cabd5ab))
+
+## [0.8.1](https://github.com/Monadical-SAS/reflector/compare/v0.8.0...v0.8.1) (2025-08-29)
+
+
+### Bug Fixes
+
+* make webhook secret/url allowing null ([#590](https://github.com/Monadical-SAS/reflector/issues/590)) ([84a3812](https://github.com/Monadical-SAS/reflector/commit/84a381220bc606231d08d6f71d4babc818fa3c75))
+
+## [0.8.0](https://github.com/Monadical-SAS/reflector/compare/v0.7.3...v0.8.0) (2025-08-29)
+
+
+### Features
+
+* **cleanup:** add automatic data retention for public instances ([#574](https://github.com/Monadical-SAS/reflector/issues/574)) ([6f0c7c1](https://github.com/Monadical-SAS/reflector/commit/6f0c7c1a5e751713366886c8e764c2009e12ba72))
+* **rooms:** add webhook for transcript completion ([#578](https://github.com/Monadical-SAS/reflector/issues/578)) ([88ed7cf](https://github.com/Monadical-SAS/reflector/commit/88ed7cfa7804794b9b54cad4c3facc8a98cf85fd))
+
+
+### Bug Fixes
+
+* file pipeline status reporting and websocket updates ([#589](https://github.com/Monadical-SAS/reflector/issues/589)) ([9dfd769](https://github.com/Monadical-SAS/reflector/commit/9dfd76996f851cc52be54feea078adbc0816dc57))
+* Igor/evaluation ([#575](https://github.com/Monadical-SAS/reflector/issues/575)) ([124ce03](https://github.com/Monadical-SAS/reflector/commit/124ce03bf86044c18313d27228a25da4bc20c9c5))
+* optimize parakeet transcription batching algorithm ([#577](https://github.com/Monadical-SAS/reflector/issues/577)) ([7030e0f](https://github.com/Monadical-SAS/reflector/commit/7030e0f23649a8cf6c1eb6d5889684a41ce849ec))
+
+## [0.7.3](https://github.com/Monadical-SAS/reflector/compare/v0.7.2...v0.7.3) (2025-08-22)
+
+
+### Bug Fixes
+
+* cleaned repo, and get git-leaks clean ([359280d](https://github.com/Monadical-SAS/reflector/commit/359280dd340433ba4402ed69034094884c825e67))
+* restore previous behavior on live pipeline + audio downscaler ([#561](https://github.com/Monadical-SAS/reflector/issues/561)) ([9265d20](https://github.com/Monadical-SAS/reflector/commit/9265d201b590d23c628c5f19251b70f473859043))
+
+## [0.7.2](https://github.com/Monadical-SAS/reflector/compare/v0.7.1...v0.7.2) (2025-08-21)
+
+
+### Bug Fixes
+
+* docker image not loading libgomp.so.1 for torch ([#560](https://github.com/Monadical-SAS/reflector/issues/560)) ([773fccd](https://github.com/Monadical-SAS/reflector/commit/773fccd93e887c3493abc2e4a4864dddce610177))
+* include shared rooms to search ([#558](https://github.com/Monadical-SAS/reflector/issues/558)) ([499eced](https://github.com/Monadical-SAS/reflector/commit/499eced3360b84fb3a90e1c8a3b554290d21adc2))
+
+## [0.7.1](https://github.com/Monadical-SAS/reflector/compare/v0.7.0...v0.7.1) (2025-08-21)
+
+
+### Bug Fixes
+
+* webvtt db null expectation mismatch ([#556](https://github.com/Monadical-SAS/reflector/issues/556)) ([e67ad1a](https://github.com/Monadical-SAS/reflector/commit/e67ad1a4a2054467bfeb1e0258fbac5868aaaf21))
+
+## [0.7.0](https://github.com/Monadical-SAS/reflector/compare/v0.6.1...v0.7.0) (2025-08-21)
+
+
+### Features
+
+* delete recording with transcript ([#547](https://github.com/Monadical-SAS/reflector/issues/547)) ([99cc984](https://github.com/Monadical-SAS/reflector/commit/99cc9840b3f5de01e0adfbfae93234042d706d13))
+* pipeline improvement with file processing, parakeet, silero-vad ([#540](https://github.com/Monadical-SAS/reflector/issues/540)) ([bcc29c9](https://github.com/Monadical-SAS/reflector/commit/bcc29c9e0050ae215f89d460e9d645aaf6a5e486))
+* postgresql migration and removal of sqlite in pytest ([#546](https://github.com/Monadical-SAS/reflector/issues/546)) ([cd1990f](https://github.com/Monadical-SAS/reflector/commit/cd1990f8f0fe1503ef5069512f33777a73a93d7f))
+* search backend ([#537](https://github.com/Monadical-SAS/reflector/issues/537)) ([5f9b892](https://github.com/Monadical-SAS/reflector/commit/5f9b89260c9ef7f3c921319719467df22830453f))
+* search frontend  ([#551](https://github.com/Monadical-SAS/reflector/issues/551)) ([3657242](https://github.com/Monadical-SAS/reflector/commit/365724271ca6e615e3425125a69ae2b46ce39285))
+
+
+### Bug Fixes
+
+* evaluation cli event wrap ([#536](https://github.com/Monadical-SAS/reflector/issues/536)) ([941c3db](https://github.com/Monadical-SAS/reflector/commit/941c3db0bdacc7b61fea412f3746cc5a7cb67836))
+* use structlog not logging ([#550](https://github.com/Monadical-SAS/reflector/issues/550)) ([27e2f81](https://github.com/Monadical-SAS/reflector/commit/27e2f81fda5232e53edc729d3e99c5ef03adbfe9))
+
+## [0.6.1](https://github.com/Monadical-SAS/reflector/compare/v0.6.0...v0.6.1) (2025-08-06)
+
+
+### Bug Fixes
+
+* delayed waveform loading ([#538](https://github.com/Monadical-SAS/reflector/issues/538)) ([ef64146](https://github.com/Monadical-SAS/reflector/commit/ef64146325d03f64dd9a1fe40234fb3e7e957ae2))
+
+## [0.6.0](https://github.com/Monadical-SAS/reflector/compare/v0.5.0...v0.6.0) (2025-08-05)
+
+
+### ⚠ BREAKING CHANGES
+
+* Configuration keys have changed. Update your .env file:
+    - TRANSCRIPT_MODAL_API_KEY → TRANSCRIPT_API_KEY
+    - LLM_MODAL_API_KEY → (removed, use TRANSCRIPT_API_KEY)
+    - Add DIARIZATION_API_KEY and TRANSLATE_API_KEY if using those services
+
+### Features
+
+* implement service-specific Modal API keys with auto processor pattern ([#528](https://github.com/Monadical-SAS/reflector/issues/528)) ([650befb](https://github.com/Monadical-SAS/reflector/commit/650befb291c47a1f49e94a01ab37d8fdfcd2b65d))
+* use llamaindex everywhere ([#525](https://github.com/Monadical-SAS/reflector/issues/525)) ([3141d17](https://github.com/Monadical-SAS/reflector/commit/3141d172bc4d3b3d533370c8e6e351ea762169bf))
+
+
+### Miscellaneous Chores
+
+* **main:** release 0.6.0 ([ecdbf00](https://github.com/Monadical-SAS/reflector/commit/ecdbf003ea2476c3e95fd231adaeb852f2943df0))
+
 ## [0.5.0](https://github.com/Monadical-SAS/reflector/compare/v0.4.0...v0.5.0) (2025-07-31)
 
 

CLAUDE.md

@@ -62,29 +62,28 @@ uv run python -m reflector.tools.process path/to/audio.wav
 **Setup:**
 ```bash
 # Install dependencies
-yarn install
+pnpm install
 
 # Copy configuration templates
 cp .env_template .env
-cp config-template.ts config.ts
 ```
 
 **Development:**
 ```bash
 # Start development server
-yarn dev
+pnpm dev
 
 # Generate TypeScript API client from OpenAPI spec
-yarn openapi
+pnpm openapi
 
 # Lint code
-yarn lint
+pnpm lint
 
 # Format code
-yarn format
+pnpm format
 
 # Build for production
-yarn build
+pnpm build
 ```
 
 ### Docker Compose (Full Stack)
@@ -144,7 +143,9 @@ All endpoints prefixed `/v1/`:
 **Backend** (`server/.env`):
 - `DATABASE_URL` - Database connection string
 - `REDIS_URL` - Redis broker for Celery
-- `MODAL_TOKEN_ID`, `MODAL_TOKEN_SECRET` - Modal.com GPU processing
+- `TRANSCRIPT_BACKEND=modal` + `TRANSCRIPT_MODAL_API_KEY` - Modal.com transcription
+- `DIARIZATION_BACKEND=modal` + `DIARIZATION_MODAL_API_KEY` - Modal.com diarization
+- `TRANSLATION_BACKEND=modal` + `TRANSLATION_MODAL_API_KEY` - Modal.com translation
 - `WHEREBY_API_KEY` - Video platform integration
 - `REFLECTOR_AUTH_BACKEND` - Authentication method (none, jwt)
 

README.md

@@ -1,43 +1,60 @@
 <div align="center">
+<img width="100" alt="image" src="https://github.com/user-attachments/assets/66fb367b-2c89-4516-9912-f47ac59c6a7f"/>
 
 # Reflector
 
-Reflector Audio Management and Analysis is a cutting-edge web application under development by Monadical. It utilizes AI to record meetings, providing a permanent record with transcripts, translations, and automated summaries.
+Reflector is an AI-powered audio transcription and meeting analysis platform that provides real-time transcription, speaker diarization, translation and summarization for audio content and live meetings. It works 100% with local models (whisper/parakeet, pyannote, seamless-m4t, and your local llm like phi-4).
 
-[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/pytests.yml)
+[![Tests](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/reflector/actions/workflows/test_server.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](https://opensource.org/licenses/MIT)
 </div>
-
-## Screenshots
+</div>
 <table>
   <tr>
     <td>
-      <a href="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/3a976930-56c1-47ef-8c76-55d3864309e3" />
+      <a href="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21f5597c-2930-4899-a154-f7bd61a59e97" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/bfe3bde3-08af-4426-a9a1-11ad5cd63b33" />
+      <a href="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/f6b9399a-5e51-4bae-b807-59128d0a940c" />
       </a>
     </td>
     <td>
-      <a href="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc">
-        <img width="700" alt="image" src="https://github.com/user-attachments/assets/7b60c9d0-efe4-474f-a27b-ea13bd0fabdc" />
+      <a href="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/a42ce460-c1fd-4489-a995-270516193897" />
+      </a>
+    </td>
+    <td>
+      <a href="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4">
+        <img width="700" alt="image" src="https://github.com/user-attachments/assets/21929f6d-c309-42fe-9c11-f1299e50fbd4" />
       </a>
     </td>
   </tr>
 </table>
 
+## What is Reflector?
+
+Reflector is a web application that utilizes local models to process audio content, providing:
+
+- **Real-time Transcription**: Convert speech to text using [Whisper](https://github.com/openai/whisper) (multi-language) or [Parakeet](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) (English) models
+- **Speaker Diarization**: Identify and label different speakers using [Pyannote](https://github.com/pyannote/pyannote-audio) 3.1
+- **Live Translation**: Translate audio content in real-time to many languages with [Facebook Seamless-M4T](https://github.com/facebookresearch/seamless_communication)
+- **Topic Detection & Summarization**: Extract key topics and generate concise summaries using LLMs
+- **Meeting Recording**: Create permanent records of meetings with searchable transcripts
+
+Currently we provide [modal.com](https://modal.com/) gpu template to deploy.
+
 ## Background
 
 The project architecture consists of three primary components:
 
-- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
 - **Back-End**: Python server that offers an API and data persistence, found in `server/`.
-- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations. Most reliable option is Modal deployment
+- **Front-End**: NextJS React project hosted on Vercel, located in `www/`.
+- **GPU implementation**: Providing services such as speech-to-text transcription, topic generation, automated summaries, and translations.
 
-It also uses authentik for authentication if activated, and Vercel for deployment and configuration of the front-end.
+It also uses authentik for authentication if activated.
 
 ## Contribution Guidelines
 
@@ -72,6 +89,8 @@ Note: We currently do not have instructions for Windows users.
 
 ## Installation
 
+*Note: we're working toward better installation, theses instructions are not accurate for now*
+
 ### Frontend
 
 Start with `cd www`.
@@ -79,17 +98,16 @@ Start with `cd www`.
 **Installation**
 
 ```bash
-yarn install
-cp .env_template .env
-cp config-template.ts config.ts
+pnpm install
+cp .env.example .env
 ```
 
-Then, fill in the environment variables in `.env` and the configuration in `config.ts` as needed. If you are unsure on how to proceed, ask in Zulip.
+Then, fill in the environment variables in `.env` as needed. If you are unsure on how to proceed, ask in Zulip.
 
 **Run in development mode**
 
 ```bash
-yarn dev
+pnpm dev
 ```
 
 Then (after completing server setup and starting it) open [http://localhost:3000](http://localhost:3000) to view it in the browser.
@@ -99,7 +117,7 @@ Then (after completing server setup and starting it) open [http://localhost:3000
 To generate the TypeScript files from the openapi.json file, make sure the python server is running, then run:
 
 ```bash
-yarn openapi
+pnpm openapi
 ```
 
 ### Backend
@@ -149,3 +167,34 @@ You can manually process an audio file by calling the process tool:
 ```bash
 uv run python -m reflector.tools.process path/to/audio.wav
 ```
+
+
+## Feature Flags
+
+Reflector uses environment variable-based feature flags to control application functionality. These flags allow you to enable or disable features without code changes.
+
+### Available Feature Flags
+
+| Feature Flag | Environment Variable |
+|-------------|---------------------|
+| `requireLogin` | `NEXT_PUBLIC_FEATURE_REQUIRE_LOGIN` |
+| `privacy` | `NEXT_PUBLIC_FEATURE_PRIVACY` |
+| `browse` | `NEXT_PUBLIC_FEATURE_BROWSE` |
+| `sendToZulip` | `NEXT_PUBLIC_FEATURE_SEND_TO_ZULIP` |
+| `rooms` | `NEXT_PUBLIC_FEATURE_ROOMS` |
+
+### Setting Feature Flags
+
+Feature flags are controlled via environment variables using the pattern `NEXT_PUBLIC_FEATURE_{FEATURE_NAME}` where `{FEATURE_NAME}` is the SCREAMING_SNAKE_CASE version of the feature name.
+
+**Examples:**
+```bash
+# Enable user authentication requirement
+NEXT_PUBLIC_FEATURE_REQUIRE_LOGIN=true
+
+# Disable browse functionality
+NEXT_PUBLIC_FEATURE_BROWSE=false
+
+# Enable Zulip integration
+NEXT_PUBLIC_FEATURE_SEND_TO_ZULIP=true
+```

compose.yml

@@ -6,6 +6,7 @@ services:
       - 1250:1250
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -16,6 +17,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -26,6 +28,7 @@ services:
       context: server
     volumes:
       - ./server/:/app/
+      - /app/.venv
     env_file:
       - ./server/.env
     environment:
@@ -39,11 +42,12 @@ services:
     image: node:18
     ports:
       - "3000:3000"
-    command: sh -c "yarn install && yarn dev"
+    command: sh -c "corepack enable && pnpm install && pnpm dev"
     restart: unless-stopped
     working_dir: /app
     volumes:
       - ./www:/app/
+      - /app/node_modules
     env_file:
       - ./www/.env.local
 

gpu/modal_deployments/.gitignore

@@ -0,0 +1,33 @@
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Logs
+*.log
+
+# Env and secrets
+.env
+.env.*
+*.env
+*.secret
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Modal local state (if any)
+modal_mounts/
+.modal_cache/

gpu/modal_deployments/README.md

@@ -0,0 +1,171 @@
+# Reflector GPU implementation - Transcription and LLM
+
+This repository hold an API for the GPU implementation of the Reflector API service,
+and use [Modal.com](https://modal.com)
+
+- `reflector_diarizer.py` - Diarization API
+- `reflector_transcriber.py` - Transcription API (Whisper)
+- `reflector_transcriber_parakeet.py` - Transcription API (NVIDIA Parakeet)
+- `reflector_translator.py` - Translation API
+
+## Modal.com deployment
+
+Create a modal secret, and name it `reflector-gpu`.
+It should contain an `REFLECTOR_APIKEY` environment variable with a value.
+
+The deployment is done using [Modal.com](https://modal.com) service.
+
+```
+$ modal deploy reflector_transcriber.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-web.modal.run
+
+$ modal deploy reflector_transcriber_parakeet.py
+...
+└── 🔨 Created web => https://xxxx--reflector-transcriber-parakeet-web.modal.run
+
+$ modal deploy reflector_llm.py
+...
+└── 🔨 Created web => https://xxxx--reflector-llm-web.modal.run
+```
+
+Then in your reflector api configuration `.env`, you can set these keys:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://xxxx--reflector-transcriber-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=REFLECTOR_APIKEY
+
+DIARIZATION_BACKEND=modal
+DIARIZATION_URL=https://xxxx--reflector-diarizer-web.modal.run
+DIARIZATION_MODAL_API_KEY=REFLECTOR_APIKEY
+
+TRANSLATION_BACKEND=modal
+TRANSLATION_URL=https://xxxx--reflector-translator-web.modal.run
+TRANSLATION_MODAL_API_KEY=REFLECTOR_APIKEY
+```
+
+## API
+
+Authentication must be passed with the `Authorization` header, using the `bearer` scheme.
+
+```
+Authorization: bearer <REFLECTOR_APIKEY>
+```
+
+### LLM
+
+`POST /llm`
+
+**request**
+```
+{
+    "prompt": "xxx"
+}
+```
+
+**response**
+```
+{
+    "text": "xxx completed"
+}
+```
+
+### Transcription
+
+#### Parakeet Transcriber (`reflector_transcriber_parakeet.py`)
+
+NVIDIA Parakeet is a state-of-the-art ASR model optimized for real-time transcription with superior word-level timestamps.
+
+**GPU Configuration:**
+- **A10G GPU** - Used for `/v1/audio/transcriptions` endpoint (small files, live transcription)
+  - Higher concurrency (max_inputs=10)
+  - Optimized for multiple small audio files
+  - Supports batch processing for efficiency
+
+- **L40S GPU** - Used for `/v1/audio/transcriptions-from-url` endpoint (large files)
+  - Lower concurrency but more powerful processing
+  - Optimized for single large audio files
+  - VAD-based chunking for long-form audio
+
+##### `/v1/audio/transcriptions` - Small file transcription
+
+**request** (multipart/form-data)
+- `file` or `files[]` - audio file(s) to transcribe
+- `model` - model name (default: `nvidia/parakeet-tdt-0.6b-v2`)
+- `language` - language code (default: `en`)
+- `batch` - whether to use batch processing for multiple files (default: `true`)
+
+**response**
+```json
+{
+    "text": "transcribed text",
+    "words": [
+        {"word": "hello", "start": 0.0, "end": 0.5},
+        {"word": "world", "start": 0.5, "end": 1.0}
+    ],
+    "filename": "audio.mp3"
+}
+```
+
+For multiple files with batch=true:
+```json
+{
+    "results": [
+        {
+            "filename": "audio1.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        },
+        {
+            "filename": "audio2.mp3",
+            "text": "transcribed text",
+            "words": [...]
+        }
+    ]
+}
+```
+
+##### `/v1/audio/transcriptions-from-url` - Large file transcription
+
+**request** (application/json)
+```json
+{
+    "audio_file_url": "https://example.com/audio.mp3",
+    "model": "nvidia/parakeet-tdt-0.6b-v2",
+    "language": "en",
+    "timestamp_offset": 0.0
+}
+```
+
+**response**
+```json
+{
+    "text": "transcribed text from large file",
+    "words": [
+        {"word": "hello", "start": 0.0, "end": 0.5},
+        {"word": "world", "start": 0.5, "end": 1.0}
+    ]
+}
+```
+
+**Supported file types:** mp3, mp4, mpeg, mpga, m4a, wav, webm
+
+#### Whisper Transcriber (`reflector_transcriber.py`)
+
+`POST /transcribe`
+
+**request** (multipart/form-data)
+
+- `file` - audio file
+- `language` - language code (e.g. `en`)
+
+**response**
+```
+{
+    "text": "xxx",
+    "words": [
+        {"text": "xxx", "start": 0.0, "end": 1.0}
+    ]
+}
+```

gpu/modal_deployments/reflector_diarizer.py

@@ -0,0 +1,253 @@
+"""
+Reflector GPU backend - diarizer
+===================================
+"""
+
+import os
+import uuid
+from typing import Mapping, NewType
+from urllib.parse import urlparse
+
+import modal
+
+PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
+MODEL_DIR = "/root/diarization_models"
+UPLOADS_PATH = "/uploads"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+
+DiarizerUniqFilename = NewType("DiarizerUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App(name="reflector-diarizer")
+
+# Volume for temporary file uploads
+upload_volume = modal.Volume.from_name("diarizer-uploads", create_if_missing=True)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
+    parsed_url = urlparse(url)
+    url_path = parsed_url.path
+
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return AudioFileExtension(ext)
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return AudioFileExtension("mp3")
+    if "audio/wav" in content_type:
+        return AudioFileExtension("wav")
+    if "audio/mp4" in content_type:
+        return AudioFileExtension("mp4")
+
+    raise ValueError(
+        f"Unsupported audio format for URL: {url}. "
+        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+    )
+
+
+def download_audio_to_volume(
+    audio_file_url: str,
+) -> tuple[DiarizerUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    print(f"Checking audio file at: {audio_file_url}")
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    print(f"Downloading audio file from: {audio_file_url}")
+    response = requests.get(audio_file_url, allow_redirects=True)
+
+    if response.status_code != 200:
+        print(f"Download failed with status {response.status_code}: {response.text}")
+        raise HTTPException(
+            status_code=response.status_code,
+            detail=f"Failed to download audio file: {response.status_code}",
+        )
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = DiarizerUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    print(f"Writing file to: {file_path} (size: {len(response.content)} bytes)")
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    print(f"File saved as: {unique_filename}")
+    return unique_filename, audio_suffix
+
+
+def migrate_cache_llm():
+    """
+    XXX The cache for model files in Transformers v4.22.0 has been updated.
+    Migrating your old cache. This is a one-time only operation. You can
+    interrupt this and resume the migration later on by calling
+    `transformers.utils.move_cache()`.
+    """
+    from transformers.utils.hub import move_cache
+
+    print("Moving LLM cache")
+    move_cache(cache_dir=MODEL_DIR, new_cache_dir=MODEL_DIR)
+    print("LLM cache moved")
+
+
+def download_pyannote_audio():
+    from pyannote.audio import Pipeline
+
+    Pipeline.from_pretrained(
+        PYANNOTE_MODEL_NAME,
+        cache_dir=MODEL_DIR,
+        use_auth_token=os.environ["HF_TOKEN"],
+    )
+
+
+diarizer_image = (
+    modal.Image.debian_slim(python_version="3.10.8")
+    .pip_install(
+        "pyannote.audio==3.1.0",
+        "requests",
+        "onnx",
+        "torchaudio",
+        "onnxruntime-gpu",
+        "torch==2.0.0",
+        "transformers==4.34.0",
+        "sentencepiece",
+        "protobuf",
+        "numpy",
+        "huggingface_hub",
+        "hf-transfer",
+    )
+    .run_function(
+        download_pyannote_audio,
+        secrets=[modal.Secret.from_name("hf_token")],
+    )
+    .run_function(migrate_cache_llm)
+    .env(
+        {
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
+            )
+        }
+    )
+)
+
+
+@app.cls(
+    gpu="A100",
+    timeout=60 * 30,
+    image=diarizer_image,
+    volumes={UPLOADS_PATH: upload_volume},
+    enable_memory_snapshot=True,
+    experimental_options={"enable_gpu_snapshot": True},
+    secrets=[
+        modal.Secret.from_name("hf_token"),
+    ],
+)
+@modal.concurrent(max_inputs=1)
+class Diarizer:
+    @modal.enter(snap=True)
+    def enter(self):
+        import torch
+        from pyannote.audio import Pipeline
+
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        print(f"Using device: {self.device}")
+        self.diarization_pipeline = Pipeline.from_pretrained(
+            PYANNOTE_MODEL_NAME,
+            cache_dir=MODEL_DIR,
+            use_auth_token=os.environ["HF_TOKEN"],
+        )
+        self.diarization_pipeline.to(torch.device(self.device))
+
+    @modal.method()
+    def diarize(self, filename: str, timestamp: float = 0.0):
+        import torchaudio
+
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        print(f"Diarizing audio from: {file_path}")
+        waveform, sample_rate = torchaudio.load(file_path)
+        diarization = self.diarization_pipeline(
+            {"waveform": waveform, "sample_rate": sample_rate}
+        )
+
+        words = []
+        for diarization_segment, _, speaker in diarization.itertracks(yield_label=True):
+            words.append(
+                {
+                    "start": round(timestamp + diarization_segment.start, 3),
+                    "end": round(timestamp + diarization_segment.end, 3),
+                    "speaker": int(speaker[-2:]),
+                }
+            )
+        print("Diarization complete")
+        return {"diarization": words}
+
+
+# -------------------------------------------------------------------
+# Web API
+# -------------------------------------------------------------------
+
+
+@app.function(
+    timeout=60 * 10,
+    scaledown_window=60 * 3,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={UPLOADS_PATH: upload_volume},
+    image=diarizer_image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import Depends, FastAPI, HTTPException, status
+    from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
+
+    diarizerstub = Diarizer()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
+            raise HTTPException(
+                status_code=status.HTTP_401_UNAUTHORIZED,
+                detail="Invalid API key",
+                headers={"WWW-Authenticate": "Bearer"},
+            )
+
+    class DiarizationResponse(BaseModel):
+        result: dict
+
+    @app.post("/diarize", dependencies=[Depends(apikey_auth)])
+    def diarize(audio_file_url: str, timestamp: float = 0.0) -> DiarizationResponse:
+        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
+
+        try:
+            func = diarizerstub.diarize.spawn(
+                filename=unique_filename, timestamp=timestamp
+            )
+            result = func.get()
+            return result
+        finally:
+            try:
+                file_path = f"{UPLOADS_PATH}/{unique_filename}"
+                print(f"Deleting file: {file_path}")
+                os.remove(file_path)
+                upload_volume.commit()
+            except Exception as e:
+                print(f"Error cleaning up {unique_filename}: {e}")
+
+    return app

gpu/modal_deployments/reflector_transcriber.py

@@ -0,0 +1,608 @@
+import os
+import sys
+import threading
+import uuid
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
+from urllib.parse import urlparse
+
+import modal
+
+MODEL_NAME = "large-v2"
+MODEL_COMPUTE_TYPE: str = "float16"
+MODEL_NUM_WORKERS: int = 1
+MINUTES = 60  # seconds
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/models"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+
+WhisperUniqFilename = NewType("WhisperUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+app = modal.App("reflector-transcriber")
+
+model_cache = modal.Volume.from_name("models", create_if_missing=True)
+upload_volume = modal.Volume.from_name("whisper-uploads", create_if_missing=True)
+
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
+def download_model():
+    from faster_whisper import download_model
+
+    model_cache.reload()
+
+    download_model(MODEL_NAME, cache_dir=CACHE_PATH)
+
+    model_cache.commit()
+
+
+image = (
+    modal.Image.debian_slim(python_version="3.12")
+    .env(
+        {
+            "HF_HUB_ENABLE_HF_TRANSFER": "1",
+            "LD_LIBRARY_PATH": (
+                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
+                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
+            ),
+        }
+    )
+    .apt_install("ffmpeg")
+    .pip_install(
+        "huggingface_hub==0.27.1",
+        "hf-transfer==0.1.9",
+        "torch==2.5.1",
+        "faster-whisper==1.1.1",
+        "fastapi==0.115.12",
+        "requests",
+        "librosa==0.10.1",
+        "numpy<2",
+        "silero-vad==5.1.0",
+    )
+    .run_function(download_model, volumes={CACHE_PATH: model_cache})
+)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
+    parsed_url = urlparse(url)
+    url_path = parsed_url.path
+
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return AudioFileExtension(ext)
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return AudioFileExtension("mp3")
+    if "audio/wav" in content_type:
+        return AudioFileExtension("wav")
+    if "audio/mp4" in content_type:
+        return AudioFileExtension("mp4")
+
+    raise ValueError(
+        f"Unsupported audio format for URL: {url}. "
+        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+    )
+
+
+def download_audio_to_volume(
+    audio_file_url: str,
+) -> tuple[WhisperUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = WhisperUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
+    """Add 0.5s of silence if audio is shorter than the silence_padding window.
+
+    Whisper does not require this strictly, but aligning behavior with Parakeet
+    avoids edge-case crashes on extremely short inputs and makes comparisons easier.
+    """
+    import numpy as np
+
+    audio_duration = len(audio_array) / sample_rate
+    if audio_duration < VAD_CONFIG["silence_padding"]:
+        silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+        silence = np.zeros(silence_samples, dtype=np.float32)
+        return np.concatenate([audio_array, silence])
+    return audio_array
+
+
+@app.cls(
+    gpu="A10G",
+    timeout=5 * MINUTES,
+    scaledown_window=5 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+@modal.concurrent(max_inputs=10)
+class TranscriberWhisperLive:
+    """Live transcriber class for small audio segments (A10G).
+
+    Mirrors the Parakeet live class API but uses Faster-Whisper under the hood.
+    """
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        print(f"Model is on device: {self.device}")
+
+    @modal.method()
+    def transcribe_segment(
+        self,
+        filename: str,
+        language: str = "en",
+    ):
+        """Transcribe a single uploaded audio file by filename."""
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    file_path,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+
+        return {"text": text, "words": words}
+
+    @modal.method()
+    def transcribe_batch(
+        self,
+        filenames: list[str],
+        language: str = "en",
+    ):
+        """Transcribe multiple uploaded audio files and return per-file results."""
+        upload_volume.reload()
+
+        results = []
+        for filename in filenames:
+            file_path = f"{UPLOADS_PATH}/{filename}"
+            if not os.path.exists(file_path):
+                raise FileNotFoundError(f"Batch file not found: {file_path}")
+
+            with self.lock:
+                with NoStdStreams():
+                    segments, _ = self.model.transcribe(
+                        file_path,
+                        language=language,
+                        beam_size=5,
+                        word_timestamps=True,
+                        vad_filter=True,
+                        vad_parameters={"min_silence_duration_ms": 500},
+                    )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start), 2),
+                    "end": round(float(w.end), 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+
+            results.append(
+                {
+                    "filename": filename,
+                    "text": text,
+                    "words": words,
+                }
+            )
+
+        return results
+
+
+@app.cls(
+    gpu="L40S",
+    timeout=15 * MINUTES,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+)
+class TranscriberWhisperFile:
+    """File transcriber for larger/longer audio, using VAD-driven batching (L40S)."""
+
+    @modal.enter()
+    def enter(self):
+        import faster_whisper
+        import torch
+        from silero_vad import load_silero_vad
+
+        self.lock = threading.Lock()
+        self.use_gpu = torch.cuda.is_available()
+        self.device = "cuda" if self.use_gpu else "cpu"
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=MODEL_COMPUTE_TYPE,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+            local_files_only=True,
+        )
+        self.vad_model = load_silero_vad(onnx=False)
+
+    @modal.method()
+    def transcribe_segment(
+        self, filename: str, timestamp_offset: float = 0.0, language: str = "en"
+    ):
+        import librosa
+        import numpy as np
+        from silero_vad import VADIterator
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLERATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[TimeSegment, None, None]:
+            """Generate speech segments as TimeSegment using Silero VAD."""
+            iterator = VADIterator(self.vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield TimeSegment(
+                        start / float(SAMPLERATE), end / float(SAMPLERATE)
+                    )
+                    start = None
+            iterator.reset_states()
+
+        upload_volume.reload()
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        audio_array, _sr = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+
+        # Batch segments up to ~30s windows by merging contiguous VAD segments
+        merged_batches: list[TimeSegment] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for segment in vad_segments(audio_array):
+            seg_start, seg_end = segment.start, segment.end
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append(TimeSegment(batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append(TimeSegment(batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+
+        for segment in merged_batches:
+            start_time, end_time = segment.start, segment.end
+            s_idx = int(start_time * SAMPLERATE)
+            e_idx = int(end_time * SAMPLERATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = pad_audio(segment, SAMPLERATE)
+
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        return {"text": " ".join(all_text), "words": all_words}
+
+
+def detect_audio_format(url: str, headers: dict) -> str:
+    from urllib.parse import urlparse
+
+    from fastapi import HTTPException
+
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_volume(audio_file_url: str) -> tuple[str, str]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+@app.function(
+    scaledown_window=60,
+    timeout=600,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    image=image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    from fastapi import (
+        Body,
+        Depends,
+        FastAPI,
+        Form,
+        HTTPException,
+        UploadFile,
+        status,
+    )
+    from fastapi.security import OAuth2PasswordBearer
+
+    transcriber_live = TranscriberWhisperLive()
+    transcriber_file = TranscriberWhisperFile()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    class TranscriptResponse(dict):
+        pass
+
+    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
+    def transcribe(
+        file: UploadFile = None,
+        files: list[UploadFile] | None = None,
+        model: str = Form(MODEL_NAME),
+        language: str = Form("en"),
+        batch: bool = Form(False),
+    ):
+        if not file and not files:
+            raise HTTPException(
+                status_code=400, detail="Either 'file' or 'files' parameter is required"
+            )
+        if batch and not files:
+            raise HTTPException(
+                status_code=400, detail="Batch transcription requires 'files'"
+            )
+
+        upload_files = [file] if file else files
+
+        uploaded_filenames: list[str] = []
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1]
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = f"{UPLOADS_PATH}/{unique_filename}"
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_filenames.append(unique_filename)
+
+        upload_volume.commit()
+
+        try:
+            if batch and len(upload_files) > 1:
+                func = transcriber_live.transcribe_batch.spawn(
+                    filenames=uploaded_filenames,
+                    language=language,
+                )
+                results = func.get()
+                return {"results": results}
+
+            results = []
+            for filename in uploaded_filenames:
+                func = transcriber_live.transcribe_segment.spawn(
+                    filename=filename,
+                    language=language,
+                )
+                result = func.get()
+                result["filename"] = filename
+                results.append(result)
+
+            return {"results": results} if len(results) > 1 else results[0]
+        finally:
+            for filename in uploaded_filenames:
+                try:
+                    file_path = f"{UPLOADS_PATH}/{filename}"
+                    os.remove(file_path)
+                except Exception:
+                    pass
+            upload_volume.commit()
+
+    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
+    def transcribe_from_url(
+        audio_file_url: str = Body(
+            ..., description="URL of the audio file to transcribe"
+        ),
+        model: str = Body(MODEL_NAME),
+        language: str = Body("en"),
+        timestamp_offset: float = Body(0.0),
+    ):
+        unique_filename, _audio_suffix = download_audio_to_volume(audio_file_url)
+        try:
+            func = transcriber_file.transcribe_segment.spawn(
+                filename=unique_filename,
+                timestamp_offset=timestamp_offset,
+                language=language,
+            )
+            result = func.get()
+            return result
+        finally:
+            try:
+                file_path = f"{UPLOADS_PATH}/{unique_filename}"
+                os.remove(file_path)
+                upload_volume.commit()
+            except Exception:
+                pass
+
+    return app
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()

gpu/modal_deployments/reflector_transcriber_parakeet.py

@@ -0,0 +1,658 @@
+import logging
+import os
+import sys
+import threading
+import uuid
+from typing import Generator, Mapping, NamedTuple, NewType, TypedDict
+from urllib.parse import urlparse
+
+import modal
+
+MODEL_NAME = "nvidia/parakeet-tdt-0.6b-v2"
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLERATE = 16000
+UPLOADS_PATH = "/uploads"
+CACHE_PATH = "/cache"
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+
+class TimeSegment(NamedTuple):
+    """Represents a time segment with start and end times."""
+
+    start: float
+    end: float
+
+
+class AudioSegment(NamedTuple):
+    """Represents an audio segment with timing and audio data."""
+
+    start: float
+    end: float
+    audio: any
+
+
+class TranscriptResult(NamedTuple):
+    """Represents a transcription result with text and word timings."""
+
+    text: str
+    words: list["WordTiming"]
+
+
+class WordTiming(TypedDict):
+    """Represents a word with its timing information."""
+
+    word: str
+    start: float
+    end: float
+
+
+app = modal.App("reflector-transcriber-parakeet")
+
+# Volume for caching model weights
+model_cache = modal.Volume.from_name("parakeet-model-cache", create_if_missing=True)
+# Volume for temporary file uploads
+upload_volume = modal.Volume.from_name("parakeet-uploads", create_if_missing=True)
+
+image = (
+    modal.Image.from_registry(
+        "nvidia/cuda:12.8.0-cudnn-devel-ubuntu22.04", add_python="3.12"
+    )
+    .env(
+        {
+            "HF_HUB_ENABLE_HF_TRANSFER": "1",
+            "HF_HOME": "/cache",
+            "DEBIAN_FRONTEND": "noninteractive",
+            "CXX": "g++",
+            "CC": "g++",
+        }
+    )
+    .apt_install("ffmpeg")
+    .pip_install(
+        "hf_transfer==0.1.9",
+        "huggingface_hub[hf-xet]==0.31.2",
+        "nemo_toolkit[asr]==2.3.0",
+        "cuda-python==12.8.0",
+        "fastapi==0.115.12",
+        "numpy<2",
+        "librosa==0.10.1",
+        "requests",
+        "silero-vad==5.1.0",
+        "torch",
+    )
+    .entrypoint([])  # silence chatty logs by container on start
+)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
+    parsed_url = urlparse(url)
+    url_path = parsed_url.path
+
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return AudioFileExtension(ext)
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return AudioFileExtension("mp3")
+    if "audio/wav" in content_type:
+        return AudioFileExtension("wav")
+    if "audio/mp4" in content_type:
+        return AudioFileExtension("mp4")
+
+    raise ValueError(
+        f"Unsupported audio format for URL: {url}. "
+        f"Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+    )
+
+
+def download_audio_to_volume(
+    audio_file_url: str,
+) -> tuple[ParakeetUniqFilename, AudioFileExtension]:
+    import requests
+    from fastapi import HTTPException
+
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = ParakeetUniqFilename(f"{uuid.uuid4()}.{audio_suffix}")
+    file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    upload_volume.commit()
+    return unique_filename, audio_suffix
+
+
+def pad_audio(audio_array, sample_rate: int = SAMPLERATE):
+    """Add 0.5 seconds of silence if audio is less than 500ms.
+
+    This is a workaround for a Parakeet bug where very short audio (<500ms) causes:
+    ValueError: `char_offsets`: [] and `processed_tokens`: [157, 834, 834, 841]
+    have to be of the same length
+
+    See: https://github.com/NVIDIA/NeMo/issues/8451
+    """
+    import numpy as np
+
+    audio_duration = len(audio_array) / sample_rate
+    if audio_duration < 0.5:
+        silence_samples = int(sample_rate * 0.5)
+        silence = np.zeros(silence_samples, dtype=np.float32)
+        return np.concatenate([audio_array, silence])
+    return audio_array
+
+
+@app.cls(
+    gpu="A10G",
+    timeout=600,
+    scaledown_window=300,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    enable_memory_snapshot=True,
+    experimental_options={"enable_gpu_snapshot": True},
+)
+@modal.concurrent(max_inputs=10)
+class TranscriberParakeetLive:
+    @modal.enter(snap=True)
+    def enter(self):
+        import nemo.collections.asr as nemo_asr
+
+        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
+
+        self.lock = threading.Lock()
+        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
+        device = next(self.model.parameters()).device
+        print(f"Model is on device: {device}")
+
+    @modal.method()
+    def transcribe_segment(
+        self,
+        filename: str,
+    ):
+        import librosa
+
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+        padded_audio = pad_audio(audio_array, sample_rate)
+
+        with self.lock:
+            with NoStdStreams():
+                (output,) = self.model.transcribe([padded_audio], timestamps=True)
+
+        text = output.text.strip()
+        words: list[WordTiming] = [
+            WordTiming(
+                # XXX the space added here is to match the output of whisper
+                # whisper add space to each words, while parakeet don't
+                word=word_info["word"] + " ",
+                start=round(word_info["start"], 2),
+                end=round(word_info["end"], 2),
+            )
+            for word_info in output.timestamp["word"]
+        ]
+
+        return {"text": text, "words": words}
+
+    @modal.method()
+    def transcribe_batch(
+        self,
+        filenames: list[str],
+    ):
+        import librosa
+
+        upload_volume.reload()
+
+        results = []
+        audio_arrays = []
+
+        # Load all audio files with padding
+        for filename in filenames:
+            file_path = f"{UPLOADS_PATH}/{filename}"
+            if not os.path.exists(file_path):
+                raise FileNotFoundError(f"Batch file not found: {file_path}")
+
+            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+            padded_audio = pad_audio(audio_array, sample_rate)
+            audio_arrays.append(padded_audio)
+
+        with self.lock:
+            with NoStdStreams():
+                outputs = self.model.transcribe(audio_arrays, timestamps=True)
+
+        # Process results for each file
+        for i, (filename, output) in enumerate(zip(filenames, outputs)):
+            text = output.text.strip()
+
+            words: list[WordTiming] = [
+                WordTiming(
+                    word=word_info["word"] + " ",
+                    start=round(word_info["start"], 2),
+                    end=round(word_info["end"], 2),
+                )
+                for word_info in output.timestamp["word"]
+            ]
+
+            results.append(
+                {
+                    "filename": filename,
+                    "text": text,
+                    "words": words,
+                }
+            )
+
+        return results
+
+
+# L40S class for file transcription (bigger files)
+@app.cls(
+    gpu="L40S",
+    timeout=900,
+    image=image,
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    enable_memory_snapshot=True,
+    experimental_options={"enable_gpu_snapshot": True},
+)
+class TranscriberParakeetFile:
+    @modal.enter(snap=True)
+    def enter(self):
+        import nemo.collections.asr as nemo_asr
+        import torch
+        from silero_vad import load_silero_vad
+
+        logging.getLogger("nemo_logger").setLevel(logging.CRITICAL)
+
+        self.model = nemo_asr.models.ASRModel.from_pretrained(model_name=MODEL_NAME)
+        device = next(self.model.parameters()).device
+        print(f"Model is on device: {device}")
+
+        torch.set_num_threads(1)
+        self.vad_model = load_silero_vad(onnx=False)
+        print("Silero VAD initialized")
+
+    @modal.method()
+    def transcribe_segment(
+        self,
+        filename: str,
+        timestamp_offset: float = 0.0,
+    ):
+        import librosa
+        import numpy as np
+        from silero_vad import VADIterator
+
+        def load_and_convert_audio(file_path):
+            audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
+            return audio_array
+
+        def vad_segment_generator(
+            audio_array,
+        ) -> Generator[TimeSegment, None, None]:
+            """Generate speech segments using VAD with start/end sample indices"""
+            vad_iterator = VADIterator(self.vad_model, sampling_rate=SAMPLERATE)
+            window_size = VAD_CONFIG["window_size"]
+            start = None
+
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+
+                speech_dict = vad_iterator(chunk)
+                if not speech_dict:
+                    continue
+
+                if "start" in speech_dict:
+                    start = speech_dict["start"]
+                    continue
+
+                if "end" in speech_dict and start is not None:
+                    end = speech_dict["end"]
+                    start_time = start / float(SAMPLERATE)
+                    end_time = end / float(SAMPLERATE)
+
+                    yield TimeSegment(start_time, end_time)
+                    start = None
+
+            vad_iterator.reset_states()
+
+        def batch_speech_segments(
+            segments: Generator[TimeSegment, None, None], max_duration: int
+        ) -> Generator[TimeSegment, None, None]:
+            """
+            Input segments:
+              [0-2] [3-5] [6-8] [10-11] [12-15] [17-19] [20-22]
+
+                                  ↓ (max_duration=10)
+
+              Output batches:
+              [0-8]           [10-19]          [20-22]
+
+            Note: silences are kept for better transcription, previous implementation was
+            passing segments separatly, but the output was less accurate.
+            """
+            batch_start_time = None
+            batch_end_time = None
+
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                if batch_start_time is None or batch_end_time is None:
+                    batch_start_time = start_time
+                    batch_end_time = end_time
+                    continue
+
+                total_duration = end_time - batch_start_time
+
+                if total_duration <= max_duration:
+                    batch_end_time = end_time
+                    continue
+
+                yield TimeSegment(batch_start_time, batch_end_time)
+                batch_start_time = start_time
+                batch_end_time = end_time
+
+            if batch_start_time is None or batch_end_time is None:
+                return
+
+            yield TimeSegment(batch_start_time, batch_end_time)
+
+        def batch_segment_to_audio_segment(
+            segments: Generator[TimeSegment, None, None],
+            audio_array,
+        ) -> Generator[AudioSegment, None, None]:
+            """Extract audio segments and apply padding for Parakeet compatibility.
+
+            Uses pad_audio to ensure segments are at least 0.5s long, preventing
+            Parakeet crashes. This padding may cause slight timing overlaps between
+            segments, which are corrected by enforce_word_timing_constraints.
+            """
+            for segment in segments:
+                start_time, end_time = segment.start, segment.end
+                start_sample = int(start_time * SAMPLERATE)
+                end_sample = int(end_time * SAMPLERATE)
+                audio_segment = audio_array[start_sample:end_sample]
+
+                padded_segment = pad_audio(audio_segment, SAMPLERATE)
+
+                yield AudioSegment(start_time, end_time, padded_segment)
+
+        def transcribe_batch(model, audio_segments: list) -> list:
+            with NoStdStreams():
+                outputs = model.transcribe(audio_segments, timestamps=True)
+            return outputs
+
+        def enforce_word_timing_constraints(
+            words: list[WordTiming],
+        ) -> list[WordTiming]:
+            """Enforce that word end times don't exceed the start time of the next word.
+
+            Due to silence padding added in batch_segment_to_audio_segment for better
+            transcription accuracy, word timings from different segments may overlap.
+            This function ensures there are no overlaps by adjusting end times.
+            """
+            if len(words) <= 1:
+                return words
+
+            enforced_words = []
+            for i, word in enumerate(words):
+                enforced_word = word.copy()
+
+                if i < len(words) - 1:
+                    next_start = words[i + 1]["start"]
+                    if enforced_word["end"] > next_start:
+                        enforced_word["end"] = next_start
+
+                enforced_words.append(enforced_word)
+
+            return enforced_words
+
+        def emit_results(
+            results: list,
+            segments_info: list[AudioSegment],
+        ) -> Generator[TranscriptResult, None, None]:
+            """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
+            for i, (output, segment) in enumerate(zip(results, segments_info)):
+                start_time, end_time = segment.start, segment.end
+                text = output.text.strip()
+                words: list[WordTiming] = [
+                    WordTiming(
+                        word=word_info["word"] + " ",
+                        start=round(
+                            word_info["start"] + start_time + timestamp_offset, 2
+                        ),
+                        end=round(word_info["end"] + start_time + timestamp_offset, 2),
+                    )
+                    for word_info in output.timestamp["word"]
+                ]
+
+                yield TranscriptResult(text, words)
+
+        upload_volume.reload()
+
+        file_path = f"{UPLOADS_PATH}/{filename}"
+        if not os.path.exists(file_path):
+            raise FileNotFoundError(f"File not found: {file_path}")
+
+        audio_array = load_and_convert_audio(file_path)
+        total_duration = len(audio_array) / float(SAMPLERATE)
+
+        all_text_parts: list[str] = []
+        all_words: list[WordTiming] = []
+
+        raw_segments = vad_segment_generator(audio_array)
+        speech_segments = batch_speech_segments(
+            raw_segments,
+            VAD_CONFIG["batch_max_duration"],
+        )
+        audio_segments = batch_segment_to_audio_segment(speech_segments, audio_array)
+
+        for batch in audio_segments:
+            audio_segment = batch.audio
+            results = transcribe_batch(self.model, [audio_segment])
+
+            for result in emit_results(
+                results,
+                [batch],
+            ):
+                if not result.text:
+                    continue
+                all_text_parts.append(result.text)
+                all_words.extend(result.words)
+
+        all_words = enforce_word_timing_constraints(all_words)
+
+        combined_text = " ".join(all_text_parts)
+        return {"text": combined_text, "words": all_words}
+
+
+@app.function(
+    scaledown_window=60,
+    timeout=600,
+    secrets=[
+        modal.Secret.from_name("reflector-gpu"),
+    ],
+    volumes={CACHE_PATH: model_cache, UPLOADS_PATH: upload_volume},
+    image=image,
+)
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
+def web():
+    import os
+    import uuid
+
+    from fastapi import (
+        Body,
+        Depends,
+        FastAPI,
+        Form,
+        HTTPException,
+        UploadFile,
+        status,
+    )
+    from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
+
+    transcriber_live = TranscriberParakeetLive()
+    transcriber_file = TranscriberParakeetFile()
+
+    app = FastAPI()
+
+    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+        if apikey == os.environ["REFLECTOR_GPU_APIKEY"]:
+            return
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED,
+            detail="Invalid API key",
+            headers={"WWW-Authenticate": "Bearer"},
+        )
+
+    class TranscriptResponse(BaseModel):
+        result: dict
+
+    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
+    def transcribe(
+        file: UploadFile = None,
+        files: list[UploadFile] | None = None,
+        model: str = Form(MODEL_NAME),
+        language: str = Form("en"),
+        batch: bool = Form(False),
+    ):
+        # Parakeet only supports English
+        if language != "en":
+            raise HTTPException(
+                status_code=400,
+                detail=f"Parakeet model only supports English. Got language='{language}'",
+            )
+        # Handle both single file and multiple files
+        if not file and not files:
+            raise HTTPException(
+                status_code=400, detail="Either 'file' or 'files' parameter is required"
+            )
+        if batch and not files:
+            raise HTTPException(
+                status_code=400, detail="Batch transcription requires 'files'"
+            )
+
+        upload_files = [file] if file else files
+
+        # Upload files to volume
+        uploaded_filenames = []
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1]
+            assert audio_suffix in SUPPORTED_FILE_EXTENSIONS
+
+            # Generate unique filename
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = f"{UPLOADS_PATH}/{unique_filename}"
+
+            print(f"Writing file to: {file_path}")
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+
+            uploaded_filenames.append(unique_filename)
+
+        upload_volume.commit()
+
+        try:
+            # Use A10G live transcriber for per-file transcription
+            if batch and len(upload_files) > 1:
+                # Use batch transcription
+                func = transcriber_live.transcribe_batch.spawn(
+                    filenames=uploaded_filenames,
+                )
+                results = func.get()
+                return {"results": results}
+
+            # Per-file transcription
+            results = []
+            for filename in uploaded_filenames:
+                func = transcriber_live.transcribe_segment.spawn(
+                    filename=filename,
+                )
+                result = func.get()
+                result["filename"] = filename
+                results.append(result)
+
+            return {"results": results} if len(results) > 1 else results[0]
+
+        finally:
+            for filename in uploaded_filenames:
+                try:
+                    file_path = f"{UPLOADS_PATH}/{filename}"
+                    print(f"Deleting file: {file_path}")
+                    os.remove(file_path)
+                except Exception as e:
+                    print(f"Error deleting {filename}: {e}")
+
+            upload_volume.commit()
+
+    @app.post("/v1/audio/transcriptions-from-url", dependencies=[Depends(apikey_auth)])
+    def transcribe_from_url(
+        audio_file_url: str = Body(
+            ..., description="URL of the audio file to transcribe"
+        ),
+        model: str = Body(MODEL_NAME),
+        language: str = Body("en", description="Language code (only 'en' supported)"),
+        timestamp_offset: float = Body(0.0),
+    ):
+        # Parakeet only supports English
+        if language != "en":
+            raise HTTPException(
+                status_code=400,
+                detail=f"Parakeet model only supports English. Got language='{language}'",
+            )
+        unique_filename, audio_suffix = download_audio_to_volume(audio_file_url)
+
+        try:
+            func = transcriber_file.transcribe_segment.spawn(
+                filename=unique_filename,
+                timestamp_offset=timestamp_offset,
+            )
+            result = func.get()
+            return result
+        finally:
+            try:
+                file_path = f"{UPLOADS_PATH}/{unique_filename}"
+                print(f"Deleting file: {file_path}")
+                os.remove(file_path)
+                upload_volume.commit()
+            except Exception as e:
+                print(f"Error cleaning up {unique_filename}: {e}")
+
+    return app
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()

gpu/self_hosted/.env.example

@@ -0,0 +1,2 @@
+REFLECTOR_GPU_APIKEY=
+HF_TOKEN=

gpu/self_hosted/.gitignore

@@ -0,0 +1,38 @@
+cache/
+
+# OS / Editor
+.DS_Store
+.vscode/
+.idea/
+
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+
+# Env and secrets
+.env
+*.env
+*.secret
+HF_TOKEN
+REFLECTOR_GPU_APIKEY
+
+# Virtual env / uv
+.venv/
+venv/
+ENV/
+uv/
+
+# Build / dist
+build/
+dist/
+.eggs/
+*.egg-info/
+
+# Coverage / test
+.pytest_cache/
+.coverage*
+htmlcov/
+
+# Logs
+*.log

gpu/self_hosted/Dockerfile

@@ -0,0 +1,46 @@
+FROM python:3.12-slim
+
+ENV PYTHONUNBUFFERED=1 \
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
+
+WORKDIR /tmp
+RUN apt-get update \
+ && apt-get install -y \
+    ffmpeg \
+    curl \
+    ca-certificates \
+    gnupg \
+    wget \
+ && apt-get clean
+# Add NVIDIA CUDA repo for Debian 12 (bookworm) and install cuDNN 9 for CUDA 12
+ADD https://developer.download.nvidia.com/compute/cuda/repos/debian12/x86_64/cuda-keyring_1.1-1_all.deb /cuda-keyring.deb
+RUN dpkg -i /cuda-keyring.deb \
+ && rm /cuda-keyring.deb \
+ && apt-get update \
+ && apt-get install -y --no-install-recommends \
+    cuda-cudart-12-6 \
+    libcublas-12-6 \
+    libcudnn9-cuda-12 \
+    libcudnn9-dev-cuda-12 \
+ && apt-get clean \
+ && rm -rf /var/lib/apt/lists/*
+ADD https://astral.sh/uv/install.sh /uv-installer.sh
+RUN sh /uv-installer.sh && rm /uv-installer.sh
+ENV PATH="/root/.local/bin/:$PATH"
+ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH"
+
+RUN mkdir -p /app
+WORKDIR /app
+COPY pyproject.toml uv.lock /app/
+
+
+COPY ./app /app/app
+COPY ./main.py /app/
+COPY ./runserver.sh /app/
+
+EXPOSE 8000
+
+CMD ["sh", "/app/runserver.sh"]
+
+

gpu/self_hosted/README.md

@@ -0,0 +1,73 @@
+# Self-hosted Model API
+
+Run transcription, translation, and diarization services compatible with Reflector's GPU Model API. Works on CPU or GPU.
+
+Environment variables
+
+- REFLECTOR_GPU_APIKEY: Optional Bearer token. If unset, auth is disabled.
+- HF_TOKEN: Optional. Required for diarization to download pyannote pipelines
+
+Requirements
+
+- FFmpeg must be installed and on PATH (used for URL-based and segmented transcription)
+- Python 3.12+
+- NVIDIA GPU optional. If available, it will be used automatically
+
+Local run
+Set env vars in self_hosted/.env file
+uv sync
+
+uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+Authentication
+
+- If REFLECTOR_GPU_APIKEY is set, include header: Authorization: Bearer <key>
+
+Endpoints
+
+- POST /v1/audio/transcriptions
+
+  - multipart/form-data
+  - fields: file (single file) OR files[] (multiple files), language, batch (true/false)
+  - response: single { text, words, filename } or { results: [ ... ] }
+
+- POST /v1/audio/transcriptions-from-url
+
+  - application/json
+  - body: { audio_file_url, language, timestamp_offset }
+  - response: { text, words }
+
+- POST /translate
+
+  - text: query parameter
+  - body (application/json): { source_language, target_language }
+  - response: { text: { <src>: original, <tgt>: translated } }
+
+- POST /diarize
+  - query parameters: audio_file_url, timestamp (optional)
+  - requires HF_TOKEN to be set (for pyannote)
+  - response: { diarization: [ { start, end, speaker } ] }
+
+OpenAPI docs
+
+- Visit /docs when the server is running
+
+Docker
+
+- Not yet provided in this directory. A Dockerfile will be added later. For now, use Local run above
+
+Conformance tests
+
+# From this directory
+
+TRANSCRIPT_URL=http://localhost:8000 \
+TRANSCRIPT_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_transcript.py
+
+TRANSLATION_URL=http://localhost:8000 \
+TRANSLATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_translation.py
+
+DIARIZATION_URL=http://localhost:8000 \
+DIARIZATION_API_KEY=dev-key \
+uv run -m pytest -m model_api --no-cov ../../server/tests/test_model_api_diarization.py

gpu/self_hosted/app/auth.py

@@ -0,0 +1,19 @@
+import os
+
+from fastapi import Depends, HTTPException, status
+from fastapi.security import OAuth2PasswordBearer
+
+oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
+
+
+def apikey_auth(apikey: str = Depends(oauth2_scheme)):
+    required_key = os.environ.get("REFLECTOR_GPU_APIKEY")
+    if not required_key:
+        return
+    if apikey == required_key:
+        return
+    raise HTTPException(
+        status_code=status.HTTP_401_UNAUTHORIZED,
+        detail="Invalid API key",
+        headers={"WWW-Authenticate": "Bearer"},
+    )

gpu/self_hosted/app/config.py

@@ -0,0 +1,12 @@
+from pathlib import Path
+
+SUPPORTED_FILE_EXTENSIONS = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
+SAMPLE_RATE = 16000
+VAD_CONFIG = {
+    "batch_max_duration": 30.0,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+# App-level paths
+UPLOADS_PATH = Path("/tmp/whisper-uploads")

gpu/self_hosted/app/factory.py

@@ -0,0 +1,30 @@
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI
+
+from .routers.diarization import router as diarization_router
+from .routers.transcription import router as transcription_router
+from .routers.translation import router as translation_router
+from .services.transcriber import WhisperService
+from .services.diarizer import PyannoteDiarizationService
+from .utils import ensure_dirs
+
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    ensure_dirs()
+    whisper_service = WhisperService()
+    whisper_service.load()
+    app.state.whisper = whisper_service
+    diarization_service = PyannoteDiarizationService()
+    diarization_service.load()
+    app.state.diarizer = diarization_service
+    yield
+
+
+def create_app() -> FastAPI:
+    app = FastAPI(lifespan=lifespan)
+    app.include_router(transcription_router)
+    app.include_router(translation_router)
+    app.include_router(diarization_router)
+    return app

gpu/self_hosted/app/routers/diarization.py

@@ -0,0 +1,30 @@
+from typing import List
+
+from fastapi import APIRouter, Depends, Request
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.diarizer import PyannoteDiarizationService
+from ..utils import download_audio_file
+
+router = APIRouter(tags=["diarization"])
+
+
+class DiarizationSegment(BaseModel):
+    start: float
+    end: float
+    speaker: int
+
+
+class DiarizationResponse(BaseModel):
+    diarization: List[DiarizationSegment]
+
+
+@router.post(
+    "/diarize", dependencies=[Depends(apikey_auth)], response_model=DiarizationResponse
+)
+def diarize(request: Request, audio_file_url: str, timestamp: float = 0.0):
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        diarizer: PyannoteDiarizationService = request.app.state.diarizer
+        return diarizer.diarize_file(file_path, timestamp=timestamp)

gpu/self_hosted/app/routers/transcription.py

@@ -0,0 +1,109 @@
+import uuid
+from typing import Optional, Union
+
+from fastapi import APIRouter, Body, Depends, Form, HTTPException, Request, UploadFile
+from pydantic import BaseModel
+from pathlib import Path
+from ..auth import apikey_auth
+from ..config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+from ..services.transcriber import MODEL_NAME
+from ..utils import cleanup_uploaded_files, download_audio_file
+
+router = APIRouter(prefix="/v1/audio", tags=["transcription"])
+
+
+class WordTiming(BaseModel):
+    word: str
+    start: float
+    end: float
+
+
+class TranscriptResult(BaseModel):
+    text: str
+    words: list[WordTiming]
+    filename: Optional[str] = None
+
+
+class TranscriptBatchResponse(BaseModel):
+    results: list[TranscriptResult]
+
+
+@router.post(
+    "/transcriptions",
+    dependencies=[Depends(apikey_auth)],
+    response_model=Union[TranscriptResult, TranscriptBatchResponse],
+)
+def transcribe(
+    request: Request,
+    file: UploadFile = None,
+    files: list[UploadFile] | None = None,
+    model: str = Form(MODEL_NAME),
+    language: str = Form("en"),
+    batch: bool = Form(False),
+):
+    service = request.app.state.whisper
+    if not file and not files:
+        raise HTTPException(
+            status_code=400, detail="Either 'file' or 'files' parameter is required"
+        )
+    if batch and not files:
+        raise HTTPException(
+            status_code=400, detail="Batch transcription requires 'files'"
+        )
+
+    upload_files = [file] if file else files
+
+    uploaded_paths: list[Path] = []
+    with cleanup_uploaded_files(uploaded_paths):
+        for upload_file in upload_files:
+            audio_suffix = upload_file.filename.split(".")[-1].lower()
+            if audio_suffix not in SUPPORTED_FILE_EXTENSIONS:
+                raise HTTPException(
+                    status_code=400,
+                    detail=(
+                        f"Unsupported audio format. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+                    ),
+                )
+            unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+            file_path = UPLOADS_PATH / unique_filename
+            with open(file_path, "wb") as f:
+                content = upload_file.file.read()
+                f.write(content)
+            uploaded_paths.append(file_path)
+
+        if batch and len(upload_files) > 1:
+            results = []
+            for path in uploaded_paths:
+                result = service.transcribe_file(str(path), language=language)
+                result["filename"] = path.name
+                results.append(result)
+            return {"results": results}
+
+        results = []
+        for path in uploaded_paths:
+            result = service.transcribe_file(str(path), language=language)
+            result["filename"] = path.name
+            results.append(result)
+
+        return {"results": results} if len(results) > 1 else results[0]
+
+
+@router.post(
+    "/transcriptions-from-url",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranscriptResult,
+)
+def transcribe_from_url(
+    request: Request,
+    audio_file_url: str = Body(..., description="URL of the audio file to transcribe"),
+    model: str = Body(MODEL_NAME),
+    language: str = Body("en"),
+    timestamp_offset: float = Body(0.0),
+):
+    service = request.app.state.whisper
+    with download_audio_file(audio_file_url) as (file_path, _ext):
+        file_path = str(file_path)
+        result = service.transcribe_vad_url_segment(
+            file_path=file_path, timestamp_offset=timestamp_offset, language=language
+        )
+        return result

gpu/self_hosted/app/routers/translation.py

@@ -0,0 +1,28 @@
+from typing import Dict
+
+from fastapi import APIRouter, Body, Depends
+from pydantic import BaseModel
+
+from ..auth import apikey_auth
+from ..services.translator import TextTranslatorService
+
+router = APIRouter(tags=["translation"])
+
+translator = TextTranslatorService()
+
+
+class TranslationResponse(BaseModel):
+    text: Dict[str, str]
+
+
+@router.post(
+    "/translate",
+    dependencies=[Depends(apikey_auth)],
+    response_model=TranslationResponse,
+)
+def translate(
+    text: str,
+    source_language: str = Body("en"),
+    target_language: str = Body("fr"),
+):
+    return translator.translate(text, source_language, target_language)

gpu/self_hosted/app/services/diarizer.py

@@ -0,0 +1,42 @@
+import os
+import threading
+
+import torch
+import torchaudio
+from pyannote.audio import Pipeline
+
+
+class PyannoteDiarizationService:
+    def __init__(self):
+        self._pipeline = None
+        self._device = "cpu"
+        self._lock = threading.Lock()
+
+    def load(self):
+        self._device = "cuda" if torch.cuda.is_available() else "cpu"
+        self._pipeline = Pipeline.from_pretrained(
+            "pyannote/speaker-diarization-3.1",
+            use_auth_token=os.environ.get("HF_TOKEN"),
+        )
+        self._pipeline.to(torch.device(self._device))
+
+    def diarize_file(self, file_path: str, timestamp: float = 0.0) -> dict:
+        if self._pipeline is None:
+            self.load()
+        waveform, sample_rate = torchaudio.load(file_path)
+        with self._lock:
+            diarization = self._pipeline(
+                {"waveform": waveform, "sample_rate": sample_rate}
+            )
+        words = []
+        for diarization_segment, _, speaker in diarization.itertracks(yield_label=True):
+            words.append(
+                {
+                    "start": round(timestamp + diarization_segment.start, 3),
+                    "end": round(timestamp + diarization_segment.end, 3),
+                    "speaker": int(speaker[-2:])
+                    if speaker and speaker[-2:].isdigit()
+                    else 0,
+                }
+            )
+        return {"diarization": words}

gpu/self_hosted/app/services/transcriber.py

@@ -0,0 +1,208 @@
+import os
+import shutil
+import subprocess
+import threading
+from typing import Generator
+
+import faster_whisper
+import librosa
+import numpy as np
+import torch
+from fastapi import HTTPException
+from silero_vad import VADIterator, load_silero_vad
+
+from ..config import SAMPLE_RATE, VAD_CONFIG
+
+# Whisper configuration (service-local defaults)
+MODEL_NAME = "large-v2"
+# None delegates compute type to runtime: float16 on CUDA, int8 on CPU
+MODEL_COMPUTE_TYPE = None
+MODEL_NUM_WORKERS = 1
+CACHE_PATH = os.path.join(os.path.expanduser("~"), ".cache", "reflector-whisper")
+from ..utils import NoStdStreams
+
+
+class WhisperService:
+    def __init__(self):
+        self.model = None
+        self.device = "cpu"
+        self.lock = threading.Lock()
+
+    def load(self):
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        compute_type = MODEL_COMPUTE_TYPE or (
+            "float16" if self.device == "cuda" else "int8"
+        )
+        self.model = faster_whisper.WhisperModel(
+            MODEL_NAME,
+            device=self.device,
+            compute_type=compute_type,
+            num_workers=MODEL_NUM_WORKERS,
+            download_root=CACHE_PATH,
+        )
+
+    def pad_audio(self, audio_array, sample_rate: int = SAMPLE_RATE):
+        audio_duration = len(audio_array) / sample_rate
+        if audio_duration < VAD_CONFIG["silence_padding"]:
+            silence_samples = int(sample_rate * VAD_CONFIG["silence_padding"])
+            silence = np.zeros(silence_samples, dtype=np.float32)
+            return np.concatenate([audio_array, silence])
+        return audio_array
+
+    def enforce_word_timing_constraints(self, words: list[dict]) -> list[dict]:
+        if len(words) <= 1:
+            return words
+        enforced: list[dict] = []
+        for i, word in enumerate(words):
+            current = dict(word)
+            if i < len(words) - 1:
+                next_start = words[i + 1]["start"]
+                if current["end"] > next_start:
+                    current["end"] = next_start
+            enforced.append(current)
+        return enforced
+
+    def transcribe_file(self, file_path: str, language: str = "en") -> dict:
+        input_for_model: str | "object" = file_path
+        try:
+            audio_array, _sample_rate = librosa.load(
+                file_path, sr=SAMPLE_RATE, mono=True
+            )
+            if len(audio_array) / float(SAMPLE_RATE) < VAD_CONFIG["silence_padding"]:
+                input_for_model = self.pad_audio(audio_array, SAMPLE_RATE)
+        except Exception:
+            pass
+
+        with self.lock:
+            with NoStdStreams():
+                segments, _ = self.model.transcribe(
+                    input_for_model,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+
+        segments = list(segments)
+        text = "".join(segment.text for segment in segments).strip()
+        words = [
+            {
+                "word": word.word,
+                "start": round(float(word.start), 2),
+                "end": round(float(word.end), 2),
+            }
+            for segment in segments
+            for word in segment.words
+        ]
+        words = self.enforce_word_timing_constraints(words)
+        return {"text": text, "words": words}
+
+    def transcribe_vad_url_segment(
+        self, file_path: str, timestamp_offset: float = 0.0, language: str = "en"
+    ) -> dict:
+        def load_audio_via_ffmpeg(input_path: str, sample_rate: int) -> np.ndarray:
+            ffmpeg_bin = shutil.which("ffmpeg") or "ffmpeg"
+            cmd = [
+                ffmpeg_bin,
+                "-nostdin",
+                "-threads",
+                "1",
+                "-i",
+                input_path,
+                "-f",
+                "f32le",
+                "-acodec",
+                "pcm_f32le",
+                "-ac",
+                "1",
+                "-ar",
+                str(sample_rate),
+                "pipe:1",
+            ]
+            try:
+                proc = subprocess.run(
+                    cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True
+                )
+            except Exception as e:
+                raise HTTPException(status_code=400, detail=f"ffmpeg failed: {e}")
+            audio = np.frombuffer(proc.stdout, dtype=np.float32)
+            return audio
+
+        def vad_segments(
+            audio_array,
+            sample_rate: int = SAMPLE_RATE,
+            window_size: int = VAD_CONFIG["window_size"],
+        ) -> Generator[tuple[float, float], None, None]:
+            vad_model = load_silero_vad(onnx=False)
+            iterator = VADIterator(vad_model, sampling_rate=sample_rate)
+            start = None
+            for i in range(0, len(audio_array), window_size):
+                chunk = audio_array[i : i + window_size]
+                if len(chunk) < window_size:
+                    chunk = np.pad(
+                        chunk, (0, window_size - len(chunk)), mode="constant"
+                    )
+                speech = iterator(chunk)
+                if not speech:
+                    continue
+                if "start" in speech:
+                    start = speech["start"]
+                    continue
+                if "end" in speech and start is not None:
+                    end = speech["end"]
+                    yield (start / float(SAMPLE_RATE), end / float(SAMPLE_RATE))
+                    start = None
+            iterator.reset_states()
+
+        audio_array = load_audio_via_ffmpeg(file_path, SAMPLE_RATE)
+
+        merged_batches: list[tuple[float, float]] = []
+        batch_start = None
+        batch_end = None
+        max_duration = VAD_CONFIG["batch_max_duration"]
+        for seg_start, seg_end in vad_segments(audio_array):
+            if batch_start is None:
+                batch_start, batch_end = seg_start, seg_end
+                continue
+            if seg_end - batch_start <= max_duration:
+                batch_end = seg_end
+            else:
+                merged_batches.append((batch_start, batch_end))
+                batch_start, batch_end = seg_start, seg_end
+        if batch_start is not None and batch_end is not None:
+            merged_batches.append((batch_start, batch_end))
+
+        all_text = []
+        all_words = []
+        for start_time, end_time in merged_batches:
+            s_idx = int(start_time * SAMPLE_RATE)
+            e_idx = int(end_time * SAMPLE_RATE)
+            segment = audio_array[s_idx:e_idx]
+            segment = self.pad_audio(segment, SAMPLE_RATE)
+            with self.lock:
+                segments, _ = self.model.transcribe(
+                    segment,
+                    language=language,
+                    beam_size=5,
+                    word_timestamps=True,
+                    vad_filter=True,
+                    vad_parameters={"min_silence_duration_ms": 500},
+                )
+            segments = list(segments)
+            text = "".join(seg.text for seg in segments).strip()
+            words = [
+                {
+                    "word": w.word,
+                    "start": round(float(w.start) + start_time + timestamp_offset, 2),
+                    "end": round(float(w.end) + start_time + timestamp_offset, 2),
+                }
+                for seg in segments
+                for w in seg.words
+            ]
+            if text:
+                all_text.append(text)
+            all_words.extend(words)
+
+        all_words = self.enforce_word_timing_constraints(all_words)
+        return {"text": " ".join(all_text), "words": all_words}

gpu/self_hosted/app/services/translator.py

@@ -0,0 +1,44 @@
+import threading
+
+from transformers import MarianMTModel, MarianTokenizer, pipeline
+
+
+class TextTranslatorService:
+    """Simple text-to-text translator using HuggingFace MarianMT models.
+
+    This mirrors the modal translator API shape but uses text translation only.
+    """
+
+    def __init__(self):
+        self._pipeline = None
+        self._lock = threading.Lock()
+
+    def load(self, source_language: str = "en", target_language: str = "fr"):
+        # Pick a default MarianMT model pair if available; fall back to Helsinki-NLP en->fr
+        model_name = self._resolve_model_name(source_language, target_language)
+        tokenizer = MarianTokenizer.from_pretrained(model_name)
+        model = MarianMTModel.from_pretrained(model_name)
+        self._pipeline = pipeline("translation", model=model, tokenizer=tokenizer)
+
+    def _resolve_model_name(self, src: str, tgt: str) -> str:
+        # Minimal mapping; extend as needed
+        pair = (src.lower(), tgt.lower())
+        mapping = {
+            ("en", "fr"): "Helsinki-NLP/opus-mt-en-fr",
+            ("fr", "en"): "Helsinki-NLP/opus-mt-fr-en",
+            ("en", "es"): "Helsinki-NLP/opus-mt-en-es",
+            ("es", "en"): "Helsinki-NLP/opus-mt-es-en",
+            ("en", "de"): "Helsinki-NLP/opus-mt-en-de",
+            ("de", "en"): "Helsinki-NLP/opus-mt-de-en",
+        }
+        return mapping.get(pair, "Helsinki-NLP/opus-mt-en-fr")
+
+    def translate(self, text: str, source_language: str, target_language: str) -> dict:
+        if self._pipeline is None:
+            self.load(source_language, target_language)
+        with self._lock:
+            results = self._pipeline(
+                text, src_lang=source_language, tgt_lang=target_language
+            )
+        translated = results[0]["translation_text"] if results else ""
+        return {"text": {source_language: text, target_language: translated}}

gpu/self_hosted/app/utils.py

@@ -0,0 +1,107 @@
+import logging
+import os
+import sys
+import uuid
+from contextlib import contextmanager
+from typing import Mapping
+from urllib.parse import urlparse
+from pathlib import Path
+
+import requests
+from fastapi import HTTPException
+
+from .config import SUPPORTED_FILE_EXTENSIONS, UPLOADS_PATH
+
+logger = logging.getLogger(__name__)
+
+
+class NoStdStreams:
+    def __init__(self):
+        self.devnull = open(os.devnull, "w")
+
+    def __enter__(self):
+        self._stdout, self._stderr = sys.stdout, sys.stderr
+        self._stdout.flush()
+        self._stderr.flush()
+        sys.stdout, sys.stderr = self.devnull, self.devnull
+
+    def __exit__(self, exc_type, exc_value, traceback):
+        sys.stdout, sys.stderr = self._stdout, self._stderr
+        self.devnull.close()
+
+
+def ensure_dirs():
+    UPLOADS_PATH.mkdir(parents=True, exist_ok=True)
+
+
+def detect_audio_format(url: str, headers: Mapping[str, str]) -> str:
+    url_path = urlparse(url).path
+    for ext in SUPPORTED_FILE_EXTENSIONS:
+        if url_path.lower().endswith(f".{ext}"):
+            return ext
+
+    content_type = headers.get("content-type", "").lower()
+    if "audio/mpeg" in content_type or "audio/mp3" in content_type:
+        return "mp3"
+    if "audio/wav" in content_type:
+        return "wav"
+    if "audio/mp4" in content_type:
+        return "mp4"
+
+    raise HTTPException(
+        status_code=400,
+        detail=(
+            f"Unsupported audio format for URL. Supported extensions: {', '.join(SUPPORTED_FILE_EXTENSIONS)}"
+        ),
+    )
+
+
+def download_audio_to_uploads(audio_file_url: str) -> tuple[Path, str]:
+    response = requests.head(audio_file_url, allow_redirects=True)
+    if response.status_code == 404:
+        raise HTTPException(status_code=404, detail="Audio file not found")
+
+    response = requests.get(audio_file_url, allow_redirects=True)
+    response.raise_for_status()
+
+    audio_suffix = detect_audio_format(audio_file_url, response.headers)
+    unique_filename = f"{uuid.uuid4()}.{audio_suffix}"
+    file_path: Path = UPLOADS_PATH / unique_filename
+
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+
+    return file_path, audio_suffix
+
+
+@contextmanager
+def download_audio_file(audio_file_url: str):
+    """Download an audio file to UPLOADS_PATH and remove it after use.
+
+    Yields (file_path: Path, audio_suffix: str).
+    """
+    file_path, audio_suffix = download_audio_to_uploads(audio_file_url)
+    try:
+        yield file_path, audio_suffix
+    finally:
+        try:
+            file_path.unlink(missing_ok=True)
+        except Exception as e:
+            logger.error("Error deleting temporary file %s: %s", file_path, e)
+
+
+@contextmanager
+def cleanup_uploaded_files(file_paths: list[Path]):
+    """Ensure provided file paths are removed after use.
+
+    The provided list can be populated inside the context; all present entries
+    at exit will be deleted.
+    """
+    try:
+        yield file_paths
+    finally:
+        for path in list(file_paths):
+            try:
+                path.unlink(missing_ok=True)
+            except Exception as e:
+                logger.error("Error deleting temporary file %s: %s", path, e)

gpu/self_hosted/compose.yml

@@ -0,0 +1,10 @@
+services:
+  reflector_gpu:
+    build:
+      context: .
+    ports:
+      - "8000:8000"
+    env_file:
+      - .env
+    volumes:
+      - ./cache:/root/.cache

gpu/self_hosted/main.py

@@ -0,0 +1,3 @@
+from app.factory import create_app
+
+app = create_app()

gpu/self_hosted/pyproject.toml

@@ -0,0 +1,19 @@
+[project]
+name = "reflector-gpu"
+version = "0.1.0"
+description = "Self-hosted GPU service for speech transcription, diarization, and translation via FastAPI."
+readme = "README.md"
+requires-python = ">=3.12"
+dependencies = [
+    "fastapi[standard]>=0.116.1",
+    "uvicorn[standard]>=0.30.0",
+    "torch>=2.3.0",
+    "faster-whisper>=1.1.0",
+    "librosa==0.10.1",
+    "numpy<2",
+    "silero-vad==5.1.0",
+    "transformers>=4.35.0",
+    "sentencepiece",
+    "pyannote.audio==3.1.0",
+    "torchaudio>=2.3.0",
+]

gpu/self_hosted/runserver.sh

@@ -0,0 +1,17 @@
+#!/bin/sh
+set -e
+
+export PATH="/root/.local/bin:$PATH"
+cd /app
+
+# Install Python dependencies at runtime (first run or when FORCE_SYNC=1)
+if [ ! -d "/app/.venv" ] || [ "$FORCE_SYNC" = "1" ]; then
+  echo "[startup] Installing Python dependencies with uv..."
+  uv sync --compile-bytecode --locked
+else
+  echo "[startup] Using existing virtual environment at /app/.venv"
+fi
+
+exec uv run uvicorn main:app --host 0.0.0.0 --port 8000
+
+

server/.gitignore

@@ -176,7 +176,8 @@ artefacts/
 audio_*.wav
 
 # ignore local database
-reflector.sqlite3
+*.sqlite3
+*.db
 data/
 
 dump.rdb

server/Dockerfile

@@ -1,7 +1,8 @@
 FROM python:3.12-slim
 
 ENV PYTHONUNBUFFERED=1 \
-    UV_LINK_MODE=copy
+    UV_LINK_MODE=copy \
+    UV_NO_CACHE=1
 
 # builder install base dependencies
 WORKDIR /tmp
@@ -13,8 +14,8 @@ ENV PATH="/root/.local/bin/:$PATH"
 # install application dependencies
 RUN mkdir -p /app
 WORKDIR /app
-COPY pyproject.toml uv.lock /app/
-RUN touch README.md && env uv sync --compile-bytecode --locked
+COPY pyproject.toml uv.lock README.md /app/
+RUN uv sync --compile-bytecode --locked
 
 # pre-download nltk packages
 RUN uv run python -c "import nltk; nltk.download('punkt_tab'); nltk.download('averaged_perceptron_tagger_eng')"
@@ -26,4 +27,15 @@ COPY migrations /app/migrations
 COPY reflector /app/reflector
 WORKDIR /app
 
+# Create symlink for libgomp if it doesn't exist (for ARM64 compatibility)
+RUN if [ "$(uname -m)" = "aarch64" ] && [ ! -f /usr/lib/libgomp.so.1 ]; then \
+    LIBGOMP_PATH=$(find /app/.venv/lib -path "*/torch.libs/libgomp*.so.*" 2>/dev/null | head -n1); \
+    if [ -n "$LIBGOMP_PATH" ]; then \
+    ln -sf "$LIBGOMP_PATH" /usr/lib/libgomp.so.1; \
+    fi \
+    fi
+
+# Pre-check just to make sure the image will not fail
+RUN uv run python -c "import silero_vad.model"
+
 CMD ["./runserver.sh"]

server/README.md

@@ -40,3 +40,5 @@ uv run python -c "from reflector.pipelines.main_live_pipeline import task_pipeli
 ```bash
 uv run python -c "from reflector.pipelines.main_live_pipeline import pipeline_post; pipeline_post(transcript_id='TRANSCRIPT_ID')"
 ```
+
+.

server/asyncio_loop_analysis.md

@@ -0,0 +1,118 @@
+# AsyncIO Event Loop Analysis for test_attendee_parsing_bug.py
+
+## Problem Summary
+The test passes but encounters an error during teardown where asyncpg tries to use a different/closed event loop, resulting in:
+- `RuntimeError: Task got Future attached to a different loop`
+- `RuntimeError: Event loop is closed`
+
+## Root Cause Analysis
+
+### 1. Multiple Event Loop Creation Points
+
+The test environment creates event loops at different scopes:
+
+1. **Session-scoped loop** (conftest.py:27-34):
+   - Created once per test session
+   - Used by session-scoped fixtures
+   - Closed after all tests complete
+
+2. **Function-scoped loop** (pytest-asyncio default):
+   - Created for each async test function
+   - This is the loop that runs the actual test
+   - Closed immediately after test completes
+
+3. **AsyncPG internal loop**:
+   - AsyncPG connections store a reference to the loop they were created with
+   - Used for connection lifecycle management
+
+### 2. Event Loop Lifecycle Mismatch
+
+The issue occurs because:
+
+1. **Session fixture creates database connection** on session-scoped loop
+2. **Test runs** on function-scoped loop (different from session loop)
+3. **During teardown**, the session fixture tries to rollback/close using the original session loop
+4. **AsyncPG connection** still references the function-scoped loop which is now closed
+5. **Conflict**: SQLAlchemy tries to use session loop, but asyncpg Future is attached to the closed function loop
+
+### 3. Configuration Issues
+
+Current pytest configuration:
+- `asyncio_mode = "auto"` in pyproject.toml
+- `asyncio_default_fixture_loop_scope=session` (shown in test output)
+- `asyncio_default_test_loop_scope=function` (shown in test output)
+
+This mismatch between fixture loop scope (session) and test loop scope (function) causes the problem.
+
+## Solutions
+
+### Option 1: Align Loop Scopes (Recommended)
+Change pytest-asyncio configuration to use consistent loop scopes:
+
+```python
+# pyproject.toml
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+asyncio_default_fixture_loop_scope = "function"  # Change from session to function
+```
+
+### Option 2: Use Function-Scoped Database Fixture
+Change the `session` fixture scope from session to function:
+
+```python
+@pytest_asyncio.fixture  # Remove scope="session"
+async def session(setup_database):
+    # ... existing code ...
+```
+
+### Option 3: Explicit Loop Management
+Ensure all async operations use the same loop:
+
+```python
+@pytest_asyncio.fixture
+async def session(setup_database, event_loop):
+    # Force using the current event loop
+    engine = create_async_engine(
+        settings.DATABASE_URL,
+        echo=False,
+        poolclass=NullPool,
+        connect_args={"loop": event_loop}  # Pass explicit loop
+    )
+    # ... rest of fixture ...
+```
+
+### Option 4: Upgrade pytest-asyncio
+The current version (1.1.0) has known issues with loop management. Consider upgrading to the latest version which has better loop scope handling.
+
+## Immediate Workaround
+
+For the test to run cleanly without the teardown error, you can:
+
+1. Add explicit cleanup in the test:
+```python
+@pytest.mark.asyncio
+async def test_attendee_parsing_bug(session):
+    # ... existing test code ...
+
+    # Explicit cleanup before fixture teardown
+    await session.commit()  # or await session.close()
+```
+
+2. Or suppress the teardown error (not recommended for production):
+```python
+@pytest.fixture
+async def session(setup_database):
+    # ... existing setup ...
+    try:
+        yield session
+        await session.rollback()
+    except RuntimeError as e:
+        if "Event loop is closed" not in str(e):
+            raise
+    finally:
+        await session.close()
+```
+
+## Recommendation
+
+The cleanest solution is to align the loop scopes by setting both fixture and test loop scopes to "function" scope. This ensures each test gets its own clean event loop and avoids cross-contamination between tests.

server/docs/data_retention.md

@@ -0,0 +1,95 @@
+# Data Retention and Cleanup
+
+## Overview
+
+For public instances of Reflector, a data retention policy is automatically enforced to delete anonymous user data after a configurable period (default: 7 days). This ensures compliance with privacy expectations and prevents unbounded storage growth.
+
+## Configuration
+
+### Environment Variables
+
+- `PUBLIC_MODE` (bool): Must be set to `true` to enable automatic cleanup
+- `PUBLIC_DATA_RETENTION_DAYS` (int): Number of days to retain anonymous data (default: 7)
+
+### What Gets Deleted
+
+When data reaches the retention period, the following items are automatically removed:
+
+1. **Transcripts** from anonymous users (where `user_id` is NULL):
+   - Database records
+   - Local files (audio.wav, audio.mp3, audio.json waveform)
+   - Storage files (cloud storage if configured)
+
+## Automatic Cleanup
+
+### Celery Beat Schedule
+
+When `PUBLIC_MODE=true`, a Celery beat task runs daily at 3 AM to clean up old data:
+
+```python
+# Automatically scheduled when PUBLIC_MODE=true
+"cleanup_old_public_data": {
+    "task": "reflector.worker.cleanup.cleanup_old_public_data",
+    "schedule": crontab(hour=3, minute=0),  # Daily at 3 AM
+}
+```
+
+### Running the Worker
+
+Ensure both Celery worker and beat scheduler are running:
+
+```bash
+# Start Celery worker
+uv run celery -A reflector.worker.app worker --loglevel=info
+
+# Start Celery beat scheduler (in another terminal)
+uv run celery -A reflector.worker.app beat
+```
+
+## Manual Cleanup
+
+For testing or manual intervention, use the cleanup tool:
+
+```bash
+# Delete data older than 7 days (default)
+uv run python -m reflector.tools.cleanup_old_data
+
+# Delete data older than 30 days
+uv run python -m reflector.tools.cleanup_old_data --days 30
+```
+
+Note: The manual tool uses the same implementation as the Celery worker task to ensure consistency.
+
+## Important Notes
+
+1. **User Data Deletion**: Only anonymous data (where `user_id` is NULL) is deleted. Authenticated user data is preserved.
+
+2. **Storage Cleanup**: The system properly cleans up both local files and cloud storage when configured.
+
+3. **Error Handling**: If individual deletions fail, the cleanup continues and logs errors. Failed deletions are reported in the task output.
+
+4. **Public Instance Only**: The automatic cleanup task only runs when `PUBLIC_MODE=true` to prevent accidental data loss in private deployments.
+
+## Testing
+
+Run the cleanup tests:
+
+```bash
+uv run pytest tests/test_cleanup.py -v
+```
+
+## Monitoring
+
+Check Celery logs for cleanup task execution:
+
+```bash
+# Look for cleanup task logs
+grep "cleanup_old_public_data" celery.log
+grep "Starting cleanup of old public data" celery.log
+```
+
+Task statistics are logged after each run:
+- Number of transcripts deleted
+- Number of meetings deleted
+- Number of orphaned recordings deleted
+- Any errors encountered

server/docs/gpu/api-transcription.md

@@ -0,0 +1,194 @@
+## Reflector GPU Transcription API (Specification)
+
+This document defines the Reflector GPU transcription API that all implementations must adhere to. Current implementations include NVIDIA Parakeet (NeMo) and Whisper (faster-whisper), both deployed on Modal.com. The API surface and response shapes are OpenAI/Whisper-compatible, so clients can switch implementations by changing only the base URL.
+
+### Base URL and Authentication
+
+- Example base URLs (Modal web endpoints):
+
+  - Parakeet: `https://<account>--reflector-transcriber-parakeet-web.modal.run`
+  - Whisper: `https://<account>--reflector-transcriber-web.modal.run`
+
+- All endpoints are served under `/v1` and require a Bearer token:
+
+```
+Authorization: Bearer <REFLECTOR_GPU_APIKEY>
+```
+
+Note: To switch implementations, deploy the desired variant and point `TRANSCRIPT_URL` to its base URL. The API is identical.
+
+### Supported file types
+
+`mp3, mp4, mpeg, mpga, m4a, wav, webm`
+
+### Models and languages
+
+- Parakeet (NVIDIA NeMo): default `nvidia/parakeet-tdt-0.6b-v2`
+  - Language support: only `en`. Other languages return HTTP 400.
+- Whisper (faster-whisper): default `large-v2` (or deployment-specific)
+  - Language support: multilingual (per Whisper model capabilities).
+
+Note: The `model` parameter is accepted by all implementations for interface parity. Some backends may treat it as informational.
+
+### Endpoints
+
+#### POST /v1/audio/transcriptions
+
+Transcribe one or more uploaded audio files.
+
+Request: multipart/form-data
+
+- `file` (File) — optional. Single file to transcribe.
+- `files` (File[]) — optional. One or more files to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`.
+  - Parakeet: only `en` is accepted; other values return HTTP 400
+  - Whisper: model-dependent; typically multilingual
+- `batch` (boolean) — optional, defaults to `false`.
+
+Notes:
+
+- Provide either `file` or `files`, not both. If neither is provided, HTTP 400.
+- `batch` requires `files`; using `batch=true` without `files` returns HTTP 400.
+- Response shape for multiple files is the same regardless of `batch`.
+- Files sent to this endpoint are processed in a single pass (no VAD/chunking). This is intended for short clips (roughly ≤ 30s; depends on GPU memory/model). For longer audio, prefer `/v1/audio/transcriptions-from-url` which supports VAD-based chunking.
+
+Responses
+
+Single file response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 0.0, "end": 0.5 },
+    { "word": "world", "start": 0.5, "end": 1.0 }
+  ],
+  "filename": "audio.mp3"
+}
+```
+
+Multiple files response:
+
+```json
+{
+  "results": [
+    {"filename": "a1.mp3", "text": "...", "words": [...]},
+    {"filename": "a2.mp3", "text": "...", "words": [...]}]
+}
+```
+
+Notes:
+
+- Word objects always include keys: `word`, `start`, `end`.
+- Some implementations may include a trailing space in `word` to match Whisper tokenization behavior; clients should trim if needed.
+
+Example curl (single file):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "file=@/path/to/audio.mp3" \
+  -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+Example curl (multiple files, batch):
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -F "files=@/path/a1.mp3" -F "files=@/path/a2.mp3" \
+  -F "batch=true" -F "language=en" \
+  "$BASE_URL/v1/audio/transcriptions"
+```
+
+#### POST /v1/audio/transcriptions-from-url
+
+Transcribe a single remote audio file by URL.
+
+Request: application/json
+
+Body parameters:
+
+- `audio_file_url` (string) — required. URL of the audio file to transcribe.
+- `model` (string) — optional. Defaults to the implementation-specific model (see above).
+- `language` (string) — optional, defaults to `en`. Parakeet only accepts `en`.
+- `timestamp_offset` (number) — optional, defaults to `0.0`. Added to each word's `start`/`end` in the response.
+
+```json
+{
+  "audio_file_url": "https://example.com/audio.mp3",
+  "model": "nvidia/parakeet-tdt-0.6b-v2",
+  "language": "en",
+  "timestamp_offset": 0.0
+}
+```
+
+Response:
+
+```json
+{
+  "text": "transcribed text",
+  "words": [
+    { "word": "hello", "start": 10.0, "end": 10.5 },
+    { "word": "world", "start": 10.5, "end": 11.0 }
+  ]
+}
+```
+
+Notes:
+
+- `timestamp_offset` is added to each word’s `start`/`end` in the response.
+- Implementations may perform VAD-based chunking and batching for long-form audio; word timings are adjusted accordingly.
+
+Example curl:
+
+```bash
+curl -X POST \
+  -H "Authorization: Bearer $REFLECTOR_GPU_APIKEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+        "audio_file_url": "https://example.com/audio.mp3",
+        "language": "en",
+        "timestamp_offset": 0
+      }' \
+  "$BASE_URL/v1/audio/transcriptions-from-url"
+```
+
+### Error handling
+
+- 400 Bad Request
+  - Parakeet: `language` other than `en`
+  - Missing required parameters (`file`/`files` for upload; `audio_file_url` for URL endpoint)
+  - Unsupported file extension
+- 401 Unauthorized
+  - Missing or invalid Bearer token
+- 404 Not Found
+  - `audio_file_url` does not exist
+
+### Implementation details
+
+- GPUs: A10G for small-file/live, L40S for large-file URL transcription (subject to deployment)
+- VAD chunking and segment batching; word timings adjusted and overlapping ends constrained
+- Pads very short segments (< 0.5s) to avoid model crashes on some backends
+
+### Server configuration (Reflector API)
+
+Set the Reflector server to use the Modal backend and point `TRANSCRIPT_URL` to your chosen deployment:
+
+```
+TRANSCRIPT_BACKEND=modal
+TRANSCRIPT_URL=https://<account>--reflector-transcriber-parakeet-web.modal.run
+TRANSCRIPT_MODAL_API_KEY=<REFLECTOR_GPU_APIKEY>
+```
+
+### Conformance tests
+
+Use the pytest-based conformance tests to validate any new implementation (including self-hosted) against this spec:
+
+```
+TRANSCRIPT_URL=https://<your-deployment-base> \
+TRANSCRIPT_MODAL_API_KEY=your-api-key \
+uv run -m pytest -m model_api --no-cov server/tests/test_model_api_transcript.py
+```

server/docs/webhook.md

@@ -0,0 +1,212 @@
+# Reflector Webhook Documentation
+
+## Overview
+
+Reflector supports webhook notifications to notify external systems when transcript processing is completed. Webhooks can be configured per room and are triggered automatically after a transcript is successfully processed.
+
+## Configuration
+
+Webhooks are configured at the room level with two fields:
+- `webhook_url`: The HTTPS endpoint to receive webhook notifications
+- `webhook_secret`: Optional secret key for HMAC signature verification (auto-generated if not provided)
+
+## Events
+
+### `transcript.completed`
+
+Triggered when a transcript has been fully processed, including transcription, diarization, summarization, and topic detection.
+
+### `test`
+
+A test event that can be triggered manually to verify webhook configuration.
+
+## Webhook Request Format
+
+### Headers
+
+All webhook requests include the following headers:
+
+| Header | Description | Example |
+|--------|-------------|---------|
+| `Content-Type` | Always `application/json` | `application/json` |
+| `User-Agent` | Identifies Reflector as the source | `Reflector-Webhook/1.0` |
+| `X-Webhook-Event` | The event type | `transcript.completed` or `test` |
+| `X-Webhook-Retry` | Current retry attempt number | `0`, `1`, `2`... |
+| `X-Webhook-Signature` | HMAC signature (if secret configured) | `t=1735306800,v1=abc123...` |
+
+### Signature Verification
+
+If a webhook secret is configured, Reflector includes an HMAC-SHA256 signature in the `X-Webhook-Signature` header to verify the webhook authenticity.
+
+The signature format is: `t={timestamp},v1={signature}`
+
+To verify the signature:
+1. Extract the timestamp and signature from the header
+2. Create the signed payload: `{timestamp}.{request_body}`
+3. Compute HMAC-SHA256 of the signed payload using your webhook secret
+4. Compare the computed signature with the received signature
+
+Example verification (Python):
+```python
+import hmac
+import hashlib
+
+def verify_webhook_signature(payload: bytes, signature_header: str, secret: str) -> bool:
+    # Parse header: "t=1735306800,v1=abc123..."
+    parts = dict(part.split("=") for part in signature_header.split(","))
+    timestamp = parts["t"]
+    received_signature = parts["v1"]
+
+    # Create signed payload
+    signed_payload = f"{timestamp}.{payload.decode('utf-8')}"
+
+    # Compute expected signature
+    expected_signature = hmac.new(
+        secret.encode("utf-8"),
+        signed_payload.encode("utf-8"),
+        hashlib.sha256
+    ).hexdigest()
+
+    # Compare signatures
+    return hmac.compare_digest(expected_signature, received_signature)
+```
+
+## Event Payloads
+
+### `transcript.completed` Event
+
+This event includes a convenient URL for accessing the transcript:
+- `frontend_url`: Direct link to view the transcript in the web interface
+
+```json
+{
+  "event": "transcript.completed",
+  "event_id": "transcript.completed-abc-123-def-456",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "transcript": {
+    "id": "abc-123-def-456",
+    "room_id": "room-789",
+    "created_at": "2025-08-27T12:00:00Z",
+    "duration": 1800.5,
+    "title": "Q3 Product Planning Meeting",
+    "short_summary": "Team discussed Q3 product roadmap, prioritizing mobile app features and API improvements.",
+    "long_summary": "The product team met to finalize the Q3 roadmap. Key decisions included...",
+    "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone to today's meeting...",
+    "topics": [
+      {
+        "title": "Introduction and Agenda",
+        "summary": "Meeting kickoff with agenda review",
+        "timestamp": 0.0,
+        "duration": 120.0,
+        "webvtt": "WEBVTT\n\n00:00:00.000 --> 00:00:05.000\n<v Speaker 1>Welcome everyone..."
+      },
+      {
+        "title": "Mobile App Features Discussion",
+        "summary": "Team reviewed proposed mobile app features for Q3",
+        "timestamp": 120.0,
+        "duration": 600.0,
+        "webvtt": "WEBVTT\n\n00:02:00.000 --> 00:02:10.000\n<v Speaker 2>Let's talk about the mobile app..."
+      }
+    ],
+    "participants": [
+      {
+        "id": "participant-1",
+        "name": "John Doe",
+        "speaker": "Speaker 1"
+      },
+      {
+        "id": "participant-2",
+        "name": "Jane Smith",
+        "speaker": "Speaker 2"
+      }
+    ],
+    "source_language": "en",
+    "target_language": "en",
+    "status": "completed",
+    "frontend_url": "https://app.reflector.com/transcripts/abc-123-def-456"
+  },
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  }
+}
+```
+
+### `test` Event
+
+```json
+{
+  "event": "test",
+  "event_id": "test.2025-08-27T12:34:56.789012Z",
+  "timestamp": "2025-08-27T12:34:56.789012Z",
+  "message": "This is a test webhook from Reflector",
+  "room": {
+    "id": "room-789",
+    "name": "Product Team Room"
+  }
+}
+```
+
+## Retry Policy
+
+Webhooks are delivered with automatic retry logic to handle transient failures. When a webhook delivery fails due to server errors or network issues, Reflector will automatically retry the delivery multiple times over an extended period.
+
+### Retry Mechanism
+
+Reflector implements an exponential backoff strategy for webhook retries:
+
+- **Initial retry delay**: 60 seconds after the first failure
+- **Exponential backoff**: Each subsequent retry waits approximately twice as long as the previous one
+- **Maximum retry interval**: 1 hour (backoff is capped at this duration)
+- **Maximum retry attempts**: 30 attempts total
+- **Total retry duration**: Retries continue for approximately 24 hours
+
+### How Retries Work
+
+When a webhook fails, Reflector will:
+1. Wait 60 seconds, then retry (attempt #1)
+2. If it fails again, wait ~2 minutes, then retry (attempt #2)
+3. Continue doubling the wait time up to a maximum of 1 hour between attempts
+4. Keep retrying at 1-hour intervals until successful or 30 attempts are exhausted
+
+The `X-Webhook-Retry` header indicates the current retry attempt number (0 for the initial attempt, 1 for first retry, etc.), allowing your endpoint to track retry attempts.
+
+### Retry Behavior by HTTP Status Code
+
+| Status Code | Behavior |
+|-------------|----------|
+| 2xx (Success) | No retry, webhook marked as delivered |
+| 4xx (Client Error) | No retry, request is considered permanently failed |
+| 5xx (Server Error) | Automatic retry with exponential backoff |
+| Network/Timeout Error | Automatic retry with exponential backoff |
+
+**Important Notes:**
+- Webhooks timeout after 30 seconds. If your endpoint takes longer to respond, it will be considered a timeout error and retried.
+- During the retry period (~24 hours), you may receive the same webhook multiple times if your endpoint experiences intermittent failures.
+- There is no mechanism to manually retry failed webhooks after the retry period expires.
+
+## Testing Webhooks
+
+You can test your webhook configuration before processing transcripts:
+
+```http
+POST /v1/rooms/{room_id}/webhook/test
+```
+
+Response:
+```json
+{
+  "success": true,
+  "status_code": 200,
+  "message": "Webhook test successful",
+  "response_preview": "OK"
+}
+```
+
+Or in case of failure:
+```json
+{
+  "success": false,
+  "error": "Webhook request timed out (10 seconds)"
+}
+```

server/env.example

@@ -24,7 +24,6 @@ AUTH_JWT_AUDIENCE=
 ## Using serverless modal.com (require reflector-gpu-modal deployed)
 #TRANSCRIPT_BACKEND=modal
 #TRANSCRIPT_URL=https://xxxxx--reflector-transcriber-web.modal.run
-#TRANSLATE_URL=https://xxxxx--reflector-translator-web.modal.run
 #TRANSCRIPT_MODAL_API_KEY=xxxxx
 
 TRANSCRIPT_BACKEND=modal
@@ -32,11 +31,13 @@ TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
 TRANSCRIPT_MODAL_API_KEY=
 
 ## =======================================================
-## Transcription backend
+## Translation backend
 ##
 ## Only available in modal atm
 ## =======================================================
+TRANSLATION_BACKEND=modal
 TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
+#TRANSLATION_MODAL_API_KEY=xxxxx
 
 ## =======================================================
 ## LLM backend
@@ -46,38 +47,11 @@ TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
 ## llm backend implementation
 ## =======================================================
 
-## Using serverless modal.com (require reflector-gpu-modal deployed)
-LLM_BACKEND=modal
-LLM_URL=https://monadical-sas--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=
-ZEPHYR_LLM_URL=https://monadical-sas--reflector-llm-zephyr-web.modal.run
-
-
-## Using OpenAI
-#LLM_BACKEND=openai
-#LLM_OPENAI_KEY=xxx
-#LLM_OPENAI_MODEL=gpt-3.5-turbo
-
-## Using GPT4ALL
-#LLM_BACKEND=openai
-#LLM_URL=http://localhost:4891/v1/completions
-#LLM_OPENAI_MODEL="GPT4All Falcon"
-
-## Default LLM MODEL NAME
-#DEFAULT_LLM=lmsys/vicuna-13b-v1.5
-
-## Cache directory to store models
-CACHE_DIR=data
-
-## =======================================================
-## Summary LLM configuration
-## =======================================================
-
 ## Context size for summary generation (tokens)
-SUMMARY_LLM_CONTEXT_SIZE_TOKENS=16000
-SUMMARY_LLM_URL=
-SUMMARY_LLM_API_KEY=sk-
-SUMMARY_MODEL=
+# LLM_MODEL=microsoft/phi-4
+LLM_CONTEXT_WINDOW=16000
+LLM_URL=
+LLM_API_KEY=sk-
 
 ## =======================================================
 ## Diarization
@@ -86,7 +60,9 @@ SUMMARY_MODEL=
 ## To allow diarization, you need to expose expose the files to be dowloded by the pipeline
 ## =======================================================
 DIARIZATION_ENABLED=false
+DIARIZATION_BACKEND=modal
 DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
+#DIARIZATION_MODAL_API_KEY=xxxxx
 
 
 ## =======================================================

server/gpu/modal_deployments/README.md

@@ -1,81 +0,0 @@
-# Reflector GPU implementation - Transcription and LLM
-
-This repository hold an API for the GPU implementation of the Reflector API service,
-and use [Modal.com](https://modal.com)
-
-- `reflector_llm.py` - LLM API
-- `reflector_transcriber.py` - Transcription API
-
-## Modal.com deployment
-
-Create a modal secret, and name it `reflector-gpu`.
-It should contain an `REFLECTOR_APIKEY` environment variable with a value.
-
-The deployment is done using [Modal.com](https://modal.com) service.
-
-```
-$ modal deploy reflector_transcriber.py
-...
-└── 🔨 Created web => https://xxxx--reflector-transcriber-web.modal.run
-
-$ modal deploy reflector_llm.py
-...
-└── 🔨 Created web => https://xxxx--reflector-llm-web.modal.run
-```
-
-Then in your reflector api configuration `.env`, you can set theses keys:
-
-```
-TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://xxxx--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=REFLECTOR_APIKEY
-
-LLM_BACKEND=modal
-LLM_URL=https://xxxx--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=REFLECTOR_APIKEY
-```
-
-## API
-
-Authentication must be passed with the `Authorization` header, using the `bearer` scheme.
-
-```
-Authorization: bearer <REFLECTOR_APIKEY>
-```
-
-### LLM
-
-`POST /llm`
-
-**request**
-```
-{
-    "prompt": "xxx"
-}
-```
-
-**response**
-```
-{
-    "text": "xxx completed"
-}
-```
-
-### Transcription
-
-`POST /transcribe`
-
-**request** (multipart/form-data)
-
-- `file` - audio file
-- `language` - language code (e.g. `en`)
-
-**response**
-```
-{
-    "text": "xxx",
-    "words": [
-        {"text": "xxx", "start": 0.0, "end": 1.0}
-    ]
-}
-```

server/gpu/modal_deployments/reflector_diarizer.py

@@ -1,187 +0,0 @@
-"""
-Reflector GPU backend - diarizer
-===================================
-"""
-
-import os
-
-import modal.gpu
-from modal import App, Image, Secret, asgi_app, enter, method
-from pydantic import BaseModel
-
-PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
-MODEL_DIR = "/root/diarization_models"
-app = App(name="reflector-diarizer")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=MODEL_DIR, new_cache_dir=MODEL_DIR)
-    print("LLM cache moved")
-
-
-def download_pyannote_audio():
-    from pyannote.audio import Pipeline
-
-    Pipeline.from_pretrained(
-        PYANNOTE_MODEL_NAME,
-        cache_dir=MODEL_DIR,
-        use_auth_token=os.environ["HF_TOKEN"],
-    )
-
-
-diarizer_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .pip_install(
-        "pyannote.audio==3.1.0",
-        "requests",
-        "onnx",
-        "torchaudio",
-        "onnxruntime-gpu",
-        "torch==2.0.0",
-        "transformers==4.34.0",
-        "sentencepiece",
-        "protobuf",
-        "numpy",
-        "huggingface_hub",
-        "hf-transfer",
-    )
-    .run_function(
-        download_pyannote_audio, secrets=[Secret.from_name("my-huggingface-secret")]
-    )
-    .run_function(migrate_cache_llm)
-    .env(
-        {
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
-            )
-        }
-    )
-)
-
-
-@app.cls(
-    gpu=modal.gpu.A100(size="40GB"),
-    timeout=60 * 30,
-    scaledown_window=60,
-    allow_concurrent_inputs=1,
-    image=diarizer_image,
-)
-class Diarizer:
-    @enter()
-    def enter(self):
-        import torch
-        from pyannote.audio import Pipeline
-
-        self.use_gpu = torch.cuda.is_available()
-        self.device = "cuda" if self.use_gpu else "cpu"
-        self.diarization_pipeline = Pipeline.from_pretrained(
-            PYANNOTE_MODEL_NAME, cache_dir=MODEL_DIR
-        )
-        self.diarization_pipeline.to(torch.device(self.device))
-
-    @method()
-    def diarize(self, audio_data: str, audio_suffix: str, timestamp: float):
-        import tempfile
-
-        import torchaudio
-
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
-
-            print("Diarizing audio")
-            waveform, sample_rate = torchaudio.load(fp.name)
-            diarization = self.diarization_pipeline(
-                {"waveform": waveform, "sample_rate": sample_rate}
-            )
-
-            words = []
-            for diarization_segment, _, speaker in diarization.itertracks(
-                yield_label=True
-            ):
-                words.append(
-                    {
-                        "start": round(timestamp + diarization_segment.start, 3),
-                        "end": round(timestamp + diarization_segment.end, 3),
-                        "speaker": int(speaker[-2:]),
-                    }
-                )
-            print("Diarization complete")
-            return {"diarization": words}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    timeout=60 * 10,
-    scaledown_window=60 * 3,
-    allow_concurrent_inputs=40,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-    image=diarizer_image,
-)
-@asgi_app()
-def web():
-    import requests
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-
-    diarizerstub = Diarizer()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    def validate_audio_file(audio_file_url: str):
-        # Check if the audio file exists
-        response = requests.head(audio_file_url, allow_redirects=True)
-        if response.status_code == 404:
-            raise HTTPException(
-                status_code=response.status_code,
-                detail="The audio file does not exist.",
-            )
-
-    class DiarizationResponse(BaseModel):
-        result: dict
-
-    @app.post(
-        "/diarize", dependencies=[Depends(apikey_auth), Depends(validate_audio_file)]
-    )
-    def diarize(
-        audio_file_url: str, timestamp: float = 0.0
-    ) -> HTTPException | DiarizationResponse:
-        # Currently the uploaded files are in mp3 format
-        audio_suffix = "mp3"
-
-        print("Downloading audio file")
-        response = requests.get(audio_file_url, allow_redirects=True)
-        print("Audio file downloaded successfully")
-
-        func = diarizerstub.diarize.spawn(
-            audio_data=response.content, audio_suffix=audio_suffix, timestamp=timestamp
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_llm.py

@@ -1,213 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-from modal import App, Image, Secret, asgi_app, enter, exit, method
-
-# LLM
-LLM_MODEL: str = "lmsys/vicuna-13b-v1.5"
-LLM_LOW_CPU_MEM_USAGE: bool = True
-LLM_TORCH_DTYPE: str = "bfloat16"
-LLM_MAX_NEW_TOKENS: int = 300
-
-IMAGE_MODEL_DIR = "/root/llm_models"
-
-app = App(name="reflector-llm")
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(LLM_MODEL, cache_dir=IMAGE_MODEL_DIR)
-    print("LLM model downloaded")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=IMAGE_MODEL_DIR, new_cache_dir=IMAGE_MODEL_DIR)
-    print("LLM cache moved")
-
-
-llm_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .apt_install("git")
-    .pip_install(
-        "transformers",
-        "torch",
-        "sentencepiece",
-        "protobuf",
-        "jsonformer==0.12.0",
-        "accelerate==0.21.0",
-        "einops==0.6.1",
-        "hf-transfer~=0.1",
-        "huggingface_hub==0.16.4",
-    )
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .run_function(download_llm)
-    .run_function(migrate_cache_llm)
-)
-
-
-@app.cls(
-    gpu="A100",
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=15,
-    image=llm_image,
-)
-class LLM:
-    @enter()
-    def enter(self):
-        import torch
-        from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
-
-        print("Instance llm model")
-        model = AutoModelForCausalLM.from_pretrained(
-            LLM_MODEL,
-            torch_dtype=getattr(torch, LLM_TORCH_DTYPE),
-            low_cpu_mem_usage=LLM_LOW_CPU_MEM_USAGE,
-            cache_dir=IMAGE_MODEL_DIR,
-            local_files_only=True,
-        )
-
-        # JSONFormer doesn't yet support generation configs
-        print("Instance llm generation config")
-        model.config.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # generation configuration
-        gen_cfg = GenerationConfig.from_model_config(model.config)
-        gen_cfg.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # load tokenizer
-        print("Instance llm tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(
-            LLM_MODEL, cache_dir=IMAGE_MODEL_DIR, local_files_only=True
-        )
-
-        # move model to gpu
-        print("Move llm model to GPU")
-        model = model.cuda()
-
-        print("Warmup llm done")
-        self.model = model
-        self.tokenizer = tokenizer
-        self.gen_cfg = gen_cfg
-        self.GenerationConfig = GenerationConfig
-
-        self.lock = threading.Lock()
-
-    @exit()
-    def exit():
-        print("Exit llm")
-
-    @method()
-    def generate(
-        self, prompt: str, gen_schema: str | None, gen_cfg: str | None
-    ) -> dict:
-        """
-        Perform a generation action using the LLM
-        """
-        print(f"Generate {prompt=}")
-        if gen_cfg:
-            gen_cfg = self.GenerationConfig.from_dict(json.loads(gen_cfg))
-        else:
-            gen_cfg = self.gen_cfg
-
-        # If a gen_schema is given, conform to gen_schema
-        with self.lock:
-            if gen_schema:
-                import jsonformer
-
-                print(f"Schema {gen_schema=}")
-                jsonformer_llm = jsonformer.Jsonformer(
-                    model=self.model,
-                    tokenizer=self.tokenizer,
-                    json_schema=json.loads(gen_schema),
-                    prompt=prompt,
-                    max_string_token_length=gen_cfg.max_new_tokens,
-                )
-                response = jsonformer_llm()
-            else:
-                # If no gen_schema, perform prompt only generation
-
-                # tokenize prompt
-                input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(
-                    self.model.device
-                )
-                output = self.model.generate(input_ids, generation_config=gen_cfg)
-
-                # decode output
-                response = self.tokenizer.decode(
-                    output[0].cpu(), skip_special_tokens=True
-                )
-                response = response[len(prompt) :]
-        print(f"Generated {response=}")
-        return {"text": response}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    scaledown_window=60 * 10,
-    timeout=60 * 5,
-    allow_concurrent_inputs=45,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-)
-@asgi_app()
-def web():
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    llmstub = LLM()
-
-    app = FastAPI()
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    class LLMRequest(BaseModel):
-        prompt: str
-        gen_schema: Optional[dict] = None
-        gen_cfg: Optional[dict] = None
-
-    @app.post("/llm", dependencies=[Depends(apikey_auth)])
-    def llm(
-        req: LLMRequest,
-    ):
-        gen_schema = json.dumps(req.gen_schema) if req.gen_schema else None
-        gen_cfg = json.dumps(req.gen_cfg) if req.gen_cfg else None
-        func = llmstub.generate.spawn(
-            prompt=req.prompt, gen_schema=gen_schema, gen_cfg=gen_cfg
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_llm_zephyr.py

@@ -1,219 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-from modal import App, Image, Secret, asgi_app, enter, exit, method
-
-# LLM
-LLM_MODEL: str = "HuggingFaceH4/zephyr-7b-alpha"
-LLM_LOW_CPU_MEM_USAGE: bool = True
-LLM_TORCH_DTYPE: str = "bfloat16"
-LLM_MAX_NEW_TOKENS: int = 300
-
-IMAGE_MODEL_DIR = "/root/llm_models/zephyr"
-
-app = App(name="reflector-llm-zephyr")
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(LLM_MODEL, cache_dir=IMAGE_MODEL_DIR)
-    print("LLM model downloaded")
-
-
-def migrate_cache_llm():
-    """
-    XXX The cache for model files in Transformers v4.22.0 has been updated.
-    Migrating your old cache. This is a one-time only operation. You can
-    interrupt this and resume the migration later on by calling
-    `transformers.utils.move_cache()`.
-    """
-    from transformers.utils.hub import move_cache
-
-    print("Moving LLM cache")
-    move_cache(cache_dir=IMAGE_MODEL_DIR, new_cache_dir=IMAGE_MODEL_DIR)
-    print("LLM cache moved")
-
-
-llm_image = (
-    Image.debian_slim(python_version="3.10.8")
-    .apt_install("git")
-    .pip_install(
-        "transformers==4.34.0",
-        "torch",
-        "sentencepiece",
-        "protobuf",
-        "jsonformer==0.12.0",
-        "accelerate==0.21.0",
-        "einops==0.6.1",
-        "hf-transfer~=0.1",
-        "huggingface_hub==0.16.4",
-    )
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .run_function(download_llm)
-    .run_function(migrate_cache_llm)
-)
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=10,
-    image=llm_image,
-)
-class LLM:
-    @enter()
-    def enter(self):
-        import torch
-        from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
-
-        print("Instance llm model")
-        model = AutoModelForCausalLM.from_pretrained(
-            LLM_MODEL,
-            torch_dtype=getattr(torch, LLM_TORCH_DTYPE),
-            low_cpu_mem_usage=LLM_LOW_CPU_MEM_USAGE,
-            cache_dir=IMAGE_MODEL_DIR,
-            local_files_only=True,
-        )
-
-        # JSONFormer doesn't yet support generation configs
-        print("Instance llm generation config")
-        model.config.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # generation configuration
-        gen_cfg = GenerationConfig.from_model_config(model.config)
-        gen_cfg.max_new_tokens = LLM_MAX_NEW_TOKENS
-
-        # load tokenizer
-        print("Instance llm tokenizer")
-        tokenizer = AutoTokenizer.from_pretrained(
-            LLM_MODEL, cache_dir=IMAGE_MODEL_DIR, local_files_only=True
-        )
-        gen_cfg.pad_token_id = tokenizer.eos_token_id
-        gen_cfg.eos_token_id = tokenizer.eos_token_id
-        tokenizer.pad_token = tokenizer.eos_token
-        model.config.pad_token_id = tokenizer.eos_token_id
-
-        # move model to gpu
-        print("Move llm model to GPU")
-        model = model.cuda()
-
-        print("Warmup llm done")
-        self.model = model
-        self.tokenizer = tokenizer
-        self.gen_cfg = gen_cfg
-        self.GenerationConfig = GenerationConfig
-        self.lock = threading.Lock()
-
-    @exit()
-    def exit():
-        print("Exit llm")
-
-    @method()
-    def generate(
-        self, prompt: str, gen_schema: str | None, gen_cfg: str | None
-    ) -> dict:
-        """
-        Perform a generation action using the LLM
-        """
-        print(f"Generate {prompt=}")
-        if gen_cfg:
-            gen_cfg = self.GenerationConfig.from_dict(json.loads(gen_cfg))
-            gen_cfg.pad_token_id = self.tokenizer.eos_token_id
-            gen_cfg.eos_token_id = self.tokenizer.eos_token_id
-        else:
-            gen_cfg = self.gen_cfg
-
-        # If a gen_schema is given, conform to gen_schema
-        with self.lock:
-            if gen_schema:
-                import jsonformer
-
-                print(f"Schema {gen_schema=}")
-                jsonformer_llm = jsonformer.Jsonformer(
-                    model=self.model,
-                    tokenizer=self.tokenizer,
-                    json_schema=json.loads(gen_schema),
-                    prompt=prompt,
-                    max_string_token_length=gen_cfg.max_new_tokens,
-                )
-                response = jsonformer_llm()
-            else:
-                # If no gen_schema, perform prompt only generation
-
-                # tokenize prompt
-                input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(
-                    self.model.device
-                )
-                output = self.model.generate(input_ids, generation_config=gen_cfg)
-
-                # decode output
-                response = self.tokenizer.decode(
-                    output[0].cpu(), skip_special_tokens=True
-                )
-                response = response[len(prompt) :]
-                response = {"long_summary": response}
-        print(f"Generated {response=}")
-        return {"text": response}
-
-
-# -------------------------------------------------------------------
-# Web API
-# -------------------------------------------------------------------
-
-
-@app.function(
-    scaledown_window=60 * 10,
-    timeout=60 * 5,
-    allow_concurrent_inputs=30,
-    secrets=[
-        Secret.from_name("reflector-gpu"),
-    ],
-)
-@asgi_app()
-def web():
-    from fastapi import Depends, FastAPI, HTTPException, status
-    from fastapi.security import OAuth2PasswordBearer
-    from pydantic import BaseModel
-
-    llmstub = LLM()
-
-    app = FastAPI()
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    class LLMRequest(BaseModel):
-        prompt: str
-        gen_schema: Optional[dict] = None
-        gen_cfg: Optional[dict] = None
-
-    @app.post("/llm", dependencies=[Depends(apikey_auth)])
-    def llm(
-        req: LLMRequest,
-    ):
-        gen_schema = json.dumps(req.gen_schema) if req.gen_schema else None
-        gen_cfg = json.dumps(req.gen_cfg) if req.gen_cfg else None
-        func = llmstub.generate.spawn(
-            prompt=req.prompt, gen_schema=gen_schema, gen_cfg=gen_cfg
-        )
-        result = func.get()
-        return result
-
-    return app

server/gpu/modal_deployments/reflector_transcriber.py

@@ -1,161 +0,0 @@
-import os
-import tempfile
-import threading
-
-import modal
-from pydantic import BaseModel
-
-MODELS_DIR = "/models"
-
-MODEL_NAME = "large-v2"
-MODEL_COMPUTE_TYPE: str = "float16"
-MODEL_NUM_WORKERS: int = 1
-
-MINUTES = 60  # seconds
-
-volume = modal.Volume.from_name("models", create_if_missing=True)
-
-app = modal.App("reflector-transcriber")
-
-
-def download_model():
-    from faster_whisper import download_model
-
-    volume.reload()
-
-    download_model(MODEL_NAME, cache_dir=MODELS_DIR)
-
-    volume.commit()
-
-
-image = (
-    modal.Image.debian_slim(python_version="3.12")
-    .pip_install(
-        "huggingface_hub==0.27.1",
-        "hf-transfer==0.1.9",
-        "torch==2.5.1",
-        "faster-whisper==1.1.1",
-    )
-    .env(
-        {
-            "HF_HUB_ENABLE_HF_TRANSFER": "1",
-            "LD_LIBRARY_PATH": (
-                "/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
-                "/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
-            ),
-        }
-    )
-    .run_function(download_model, volumes={MODELS_DIR: volume})
-)
-
-
-@app.cls(
-    gpu="A10G",
-    timeout=5 * MINUTES,
-    scaledown_window=5 * MINUTES,
-    allow_concurrent_inputs=6,
-    image=image,
-    volumes={MODELS_DIR: volume},
-)
-class Transcriber:
-    @modal.enter()
-    def enter(self):
-        import faster_whisper
-        import torch
-
-        self.lock = threading.Lock()
-        self.use_gpu = torch.cuda.is_available()
-        self.device = "cuda" if self.use_gpu else "cpu"
-        self.model = faster_whisper.WhisperModel(
-            MODEL_NAME,
-            device=self.device,
-            compute_type=MODEL_COMPUTE_TYPE,
-            num_workers=MODEL_NUM_WORKERS,
-            download_root=MODELS_DIR,
-            local_files_only=True,
-        )
-
-    @modal.method()
-    def transcribe_segment(
-        self,
-        audio_data: str,
-        audio_suffix: str,
-        language: str,
-    ):
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
-
-            with self.lock:
-                segments, _ = self.model.transcribe(
-                    fp.name,
-                    language=language,
-                    beam_size=5,
-                    word_timestamps=True,
-                    vad_filter=True,
-                    vad_parameters={"min_silence_duration_ms": 500},
-                )
-
-            segments = list(segments)
-            text = "".join(segment.text for segment in segments)
-            words = [
-                {"word": word.word, "start": word.start, "end": word.end}
-                for segment in segments
-                for word in segment.words
-            ]
-
-            return {"text": text, "words": words}
-
-
-@app.function(
-    scaledown_window=60,
-    timeout=60,
-    allow_concurrent_inputs=40,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-    volumes={MODELS_DIR: volume},
-)
-@modal.asgi_app()
-def web():
-    from fastapi import Body, Depends, FastAPI, HTTPException, UploadFile, status
-    from fastapi.security import OAuth2PasswordBearer
-    from typing_extensions import Annotated
-
-    transcriber = Transcriber()
-
-    app = FastAPI()
-
-    oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
-
-    supported_file_types = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
-
-    def apikey_auth(apikey: str = Depends(oauth2_scheme)):
-        if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
-            raise HTTPException(
-                status_code=status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid API key",
-                headers={"WWW-Authenticate": "Bearer"},
-            )
-
-    class TranscriptResponse(BaseModel):
-        result: dict
-
-    @app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
-    def transcribe(
-        file: UploadFile,
-        model: str = "whisper-1",
-        language: Annotated[str, Body(...)] = "en",
-    ) -> TranscriptResponse:
-        audio_data = file.file.read()
-        audio_suffix = file.filename.split(".")[-1]
-        assert audio_suffix in supported_file_types
-
-        func = transcriber.transcribe_segment.spawn(
-            audio_data=audio_data,
-            audio_suffix=audio_suffix,
-            language=language,
-        )
-        result = func.get()
-        return result
-
-    return app

server/migrations/README

@@ -1 +1,3 @@
-Generic single-database configuration.
+Generic single-database configuration.
+
+Both data migrations and schema migrations must be in migrations.

server/migrations/env.py

@@ -3,7 +3,7 @@ from logging.config import fileConfig
 from alembic import context
 from sqlalchemy import engine_from_config, pool
 
-from reflector.db import metadata
+from reflector.db.base import metadata
 from reflector.settings import settings
 
 # this is the Alembic Config object, which provides

server/migrations/versions/0194f65cd6d3_add_webhook_fields_to_rooms.py

@@ -0,0 +1,36 @@
+"""Add webhook fields to rooms
+
+Revision ID: 0194f65cd6d3
+Revises: 5a8907fd1d78
+Create Date: 2025-08-27 09:03:19.610995
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0194f65cd6d3"
+down_revision: Union[str, None] = "5a8907fd1d78"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("webhook_url", sa.String(), nullable=True))
+        batch_op.add_column(sa.Column("webhook_secret", sa.String(), nullable=True))
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_column("webhook_secret")
+        batch_op.drop_column("webhook_url")
+
+    # ### end Alembic commands ###

server/migrations/versions/0ab2d7ffaa16_add_long_summary_to_search_vector.py

@@ -0,0 +1,64 @@
+"""add_long_summary_to_search_vector
+
+Revision ID: 0ab2d7ffaa16
+Revises: b1c33bd09963
+Create Date: 2025-08-15 13:27:52.680211
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0ab2d7ffaa16"
+down_revision: Union[str, None] = "b1c33bd09963"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Drop the existing search vector column and index
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")
+
+    # Recreate the search vector column with long_summary included
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(long_summary, '')), 'B') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'C')
+        ) STORED
+    """)
+
+    # Recreate the GIN index for the search vector
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )
+
+
+def downgrade() -> None:
+    # Drop the updated search vector column and index
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")
+
+    # Recreate the original search vector column without long_summary
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'B')
+        ) STORED
+    """)
+
+    # Recreate the GIN index for the search vector
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )

server/migrations/versions/0bc0f3ff0111_add_webvtt_field_to_transcript.py

@@ -0,0 +1,25 @@
+"""add_webvtt_field_to_transcript
+
+Revision ID: 0bc0f3ff0111
+Revises: b7df9609542c
+Create Date: 2025-08-05 19:36:41.740957
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0bc0f3ff0111"
+down_revision: Union[str, None] = "b7df9609542c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.add_column("transcript", sa.Column("webvtt", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("transcript", "webvtt")

server/migrations/versions/0ce521cda2ee_remove_user_id_from_meeting_table.py

@@ -0,0 +1,36 @@
+"""remove user_id from meeting table
+
+Revision ID: 0ce521cda2ee
+Revises: 6dec9fb5b46c
+Create Date: 2025-09-10 12:40:55.688899
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "0ce521cda2ee"
+down_revision: Union[str, None] = "6dec9fb5b46c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("user_id")
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(
+            sa.Column("user_id", sa.VARCHAR(), autoincrement=False, nullable=True)
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/116b2f287eab_add_full_text_search.py

@@ -0,0 +1,46 @@
+"""add_full_text_search
+
+Revision ID: 116b2f287eab
+Revises: 0bc0f3ff0111
+Create Date: 2025-08-07 11:27:38.473517
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+revision: str = "116b2f287eab"
+down_revision: Union[str, None] = "0bc0f3ff0111"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    if conn.dialect.name != "postgresql":
+        return
+
+    op.execute("""
+        ALTER TABLE transcript ADD COLUMN search_vector_en tsvector
+        GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(webvtt, '')), 'B')
+        ) STORED
+    """)
+
+    op.create_index(
+        "idx_transcript_search_vector_en",
+        "transcript",
+        ["search_vector_en"],
+        postgresql_using="gin",
+    )
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    if conn.dialect.name != "postgresql":
+        return
+
+    op.drop_index("idx_transcript_search_vector_en", table_name="transcript")
+    op.drop_column("transcript", "search_vector_en")

server/migrations/versions/2ae3db106d4e_clean_up_orphaned_room_id_references_in_.py

@@ -0,0 +1,32 @@
+"""clean up orphaned room_id references in meeting table
+
+Revision ID: 2ae3db106d4e
+Revises: def1b5867d4c
+Create Date: 2025-09-11 10:35:15.759967
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "2ae3db106d4e"
+down_revision: Union[str, None] = "def1b5867d4c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Set room_id to NULL for meetings that reference non-existent rooms
+    op.execute("""
+        UPDATE meeting
+        SET room_id = NULL
+        WHERE room_id IS NOT NULL
+          AND room_id NOT IN (SELECT id FROM room WHERE id IS NOT NULL)
+    """)
+
+
+def downgrade() -> None:
+    # Cannot restore orphaned references - no operation needed
+    pass

server/migrations/versions/38a927dcb099_rename_back_text_to_transcript.py

@@ -28,7 +28,7 @@ def upgrade() -> None:
     transcript = table("transcript", column("id", sa.String), column("topics", sa.JSON))
 
     # Select all rows from the transcript table
-    results = bind.execute(select([transcript.c.id, transcript.c.topics]))
+    results = bind.execute(select(transcript.c.id, transcript.c.topics))
 
     for row in results:
         transcript_id = row["id"]
@@ -58,7 +58,7 @@ def downgrade() -> None:
     transcript = table("transcript", column("id", sa.String), column("topics", sa.JSON))
 
     # Select all rows from the transcript table
-    results = bind.execute(select([transcript.c.id, transcript.c.topics]))
+    results = bind.execute(select(transcript.c.id, transcript.c.topics))
 
     for row in results:
         transcript_id = row["id"]

server/migrations/versions/4814901632bc_fix_duration.py

@@ -36,9 +36,7 @@ def upgrade() -> None:
 
     # select only the one with duration = 0
     results = bind.execute(
-        select([transcript.c.id, transcript.c.duration]).where(
-            transcript.c.duration == 0
-        )
+        select(transcript.c.id, transcript.c.duration).where(transcript.c.duration == 0)
     )
 
     data_dir = Path(settings.DATA_DIR)

server/migrations/versions/5a8907fd1d78_add_cascade_delete_to_meeting_consent_.py

@@ -0,0 +1,50 @@
+"""add cascade delete to meeting consent foreign key
+
+Revision ID: 5a8907fd1d78
+Revises: 0ab2d7ffaa16
+Create Date: 2025-08-26 17:26:50.945491
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "5a8907fd1d78"
+down_revision: Union[str, None] = "0ab2d7ffaa16"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+            ondelete="CASCADE",
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.drop_constraint(
+            batch_op.f("meeting_consent_meeting_id_fkey"), type_="foreignkey"
+        )
+        batch_op.create_foreign_key(
+            batch_op.f("meeting_consent_meeting_id_fkey"),
+            "meeting",
+            ["meeting_id"],
+            ["id"],
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/6025e9b2bef2_remove_one_active_meeting_per_room_.py

@@ -0,0 +1,53 @@
+"""remove_one_active_meeting_per_room_constraint
+
+Revision ID: 6025e9b2bef2
+Revises: 2ae3db106d4e
+Create Date: 2025-08-18 18:45:44.418392
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6025e9b2bef2"
+down_revision: Union[str, None] = "2ae3db106d4e"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove the unique constraint that prevents multiple active meetings per room
+    # This is needed to support calendar integration with overlapping meetings
+    # Check if index exists before trying to drop it
+    from alembic import context
+
+    if context.get_context().dialect.name == "postgresql":
+        conn = op.get_bind()
+        result = conn.execute(
+            sa.text(
+                "SELECT 1 FROM pg_indexes WHERE indexname = 'idx_one_active_meeting_per_room'"
+            )
+        )
+        if result.fetchone():
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+    else:
+        # For SQLite, just try to drop it
+        try:
+            op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")
+        except:
+            pass
+
+
+def downgrade() -> None:
+    # Restore the unique constraint
+    op.create_index(
+        "idx_one_active_meeting_per_room",
+        "meeting",
+        ["room_id"],
+        unique=True,
+        postgresql_where=sa.text("is_active = true"),
+        sqlite_where=sa.text("is_active = 1"),
+    )

server/migrations/versions/61882a919591_webhook_url_and_secret_null_by_default.py

@@ -0,0 +1,28 @@
+"""webhook url and secret null by default
+
+
+Revision ID: 61882a919591
+Revises: 0194f65cd6d3
+Create Date: 2025-08-29 11:46:36.738091
+
+"""
+
+from typing import Sequence, Union
+
+# revision identifiers, used by Alembic.
+revision: str = "61882a919591"
+down_revision: Union[str, None] = "0194f65cd6d3"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    pass
+    # ### end Alembic commands ###

server/migrations/versions/62dea3db63a5_add_room_options.py

@@ -32,7 +32,7 @@ def upgrade() -> None:
         sa.Column("user_id", sa.String(), nullable=True),
         sa.Column("room_id", sa.String(), nullable=True),
         sa.Column(
-            "is_locked", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "is_locked", sa.Boolean(), server_default=sa.text("false"), nullable=False
         ),
         sa.Column("room_mode", sa.String(), server_default="normal", nullable=False),
         sa.Column(
@@ -53,12 +53,15 @@ def upgrade() -> None:
         sa.Column("user_id", sa.String(), nullable=False),
         sa.Column("created_at", sa.DateTime(), nullable=False),
         sa.Column(
-            "zulip_auto_post", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "zulip_auto_post",
+            sa.Boolean(),
+            server_default=sa.text("false"),
+            nullable=False,
         ),
         sa.Column("zulip_stream", sa.String(), nullable=True),
         sa.Column("zulip_topic", sa.String(), nullable=True),
         sa.Column(
-            "is_locked", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "is_locked", sa.Boolean(), server_default=sa.text("false"), nullable=False
         ),
         sa.Column("room_mode", sa.String(), server_default="normal", nullable=False),
         sa.Column(

server/migrations/versions/6dec9fb5b46c_make_meeting_room_id_required_and_add_.py

@@ -0,0 +1,35 @@
+"""make meeting room_id required and add foreign key
+
+Revision ID: 6dec9fb5b46c
+Revises: 61882a919591
+Create Date: 2025-09-10 10:47:06.006819
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "6dec9fb5b46c"
+down_revision: Union[str, None] = "61882a919591"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.create_foreign_key(
+            None, "room", ["room_id"], ["id"], ondelete="CASCADE"
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("meeting_room_id_fkey", type_="foreignkey")
+
+    # ### end Alembic commands ###

server/migrations/versions/74b2b0236931_add_transcript_source_kind.py

@@ -20,11 +20,14 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
+    sourcekind_enum = sa.Enum("room", "live", "file", name="sourcekind")
+    sourcekind_enum.create(op.get_bind())
+
     op.add_column(
         "transcript",
         sa.Column(
             "source_kind",
-            sa.Enum("ROOM", "LIVE", "FILE", name="sourcekind"),
+            sourcekind_enum,
             nullable=True,
         ),
     )
@@ -43,6 +46,8 @@ def upgrade() -> None:
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.drop_column("transcript", "source_kind")
+    sourcekind_enum = sa.Enum(name="sourcekind")
+    sourcekind_enum.drop(op.get_bind())
 
 
 # ### end Alembic commands ###

server/migrations/versions/8120ebc75366_populate_webvtt_from_topics.py

@@ -0,0 +1,106 @@
+"""populate_webvtt_from_topics
+
+Revision ID: 8120ebc75366
+Revises: 116b2f287eab
+Create Date: 2025-08-11 19:11:01.316947
+
+"""
+
+import json
+from typing import Sequence, Union
+
+from alembic import op
+from sqlalchemy import text
+
+# revision identifiers, used by Alembic.
+revision: str = "8120ebc75366"
+down_revision: Union[str, None] = "116b2f287eab"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def topics_to_webvtt(topics):
+    """Convert topics list to WebVTT format string."""
+    if not topics:
+        return None
+
+    lines = ["WEBVTT", ""]
+
+    for topic in topics:
+        start_time = format_timestamp(topic.get("start"))
+        end_time = format_timestamp(topic.get("end"))
+        text = topic.get("text", "").strip()
+
+        if start_time and end_time and text:
+            lines.append(f"{start_time} --> {end_time}")
+            lines.append(text)
+            lines.append("")
+
+    return "\n".join(lines).strip()
+
+
+def format_timestamp(seconds):
+    """Format seconds to WebVTT timestamp format (HH:MM:SS.mmm)."""
+    if seconds is None:
+        return None
+
+    hours = int(seconds // 3600)
+    minutes = int((seconds % 3600) // 60)
+    secs = seconds % 60
+
+    return f"{hours:02d}:{minutes:02d}:{secs:06.3f}"
+
+
+def upgrade() -> None:
+    """Populate WebVTT field for all transcripts with topics."""
+
+    # Get connection
+    connection = op.get_bind()
+
+    # Query all transcripts with topics
+    result = connection.execute(
+        text("SELECT id, topics FROM transcript WHERE topics IS NOT NULL")
+    )
+
+    rows = result.fetchall()
+    print(f"Found {len(rows)} transcripts with topics")
+
+    updated_count = 0
+    error_count = 0
+
+    for row in rows:
+        transcript_id = row[0]
+        topics_data = row[1]
+
+        if not topics_data:
+            continue
+
+        try:
+            # Parse JSON if it's a string
+            if isinstance(topics_data, str):
+                topics_data = json.loads(topics_data)
+
+            # Convert topics to WebVTT format
+            webvtt_content = topics_to_webvtt(topics_data)
+
+            if webvtt_content:
+                # Update the webvtt field
+                connection.execute(
+                    text("UPDATE transcript SET webvtt = :webvtt WHERE id = :id"),
+                    {"webvtt": webvtt_content, "id": transcript_id},
+                )
+                updated_count += 1
+                print(f"✓ Updated transcript {transcript_id}")
+
+        except Exception as e:
+            error_count += 1
+            print(f"✗ Error updating transcript {transcript_id}: {e}")
+
+    print(f"\nMigration complete!")
+    print(f"  Updated: {updated_count}")
+    print(f"  Errors: {error_count}")
+
+
+def downgrade() -> None:
+    """Clear WebVTT field for all transcripts."""
+    op.execute(text("UPDATE transcript SET webvtt = NULL"))

server/migrations/versions/9920ecfe2735_rename_transcript_to_text.py

@@ -28,7 +28,7 @@ def upgrade() -> None:
     transcript = table("transcript", column("id", sa.String), column("topics", sa.JSON))
 
     # Select all rows from the transcript table
-    results = bind.execute(select([transcript.c.id, transcript.c.topics]))
+    results = bind.execute(select(transcript.c.id, transcript.c.topics))
 
     for row in results:
         transcript_id = row["id"]
@@ -58,7 +58,7 @@ def downgrade() -> None:
     transcript = table("transcript", column("id", sa.String), column("topics", sa.JSON))
 
     # Select all rows from the transcript table
-    results = bind.execute(select([transcript.c.id, transcript.c.topics]))
+    results = bind.execute(select(transcript.c.id, transcript.c.topics))
 
     for row in results:
         transcript_id = row["id"]

server/migrations/versions/99365b0cd87b_add_title_short_and_long_summary_and_.py

@@ -22,7 +22,7 @@ def upgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.execute(
         "UPDATE transcript SET events = "
-        'REPLACE(events, \'"event": "SUMMARY"\', \'"event": "LONG_SUMMARY"\');'
+        'REPLACE(events::text, \'"event": "SUMMARY"\', \'"event": "LONG_SUMMARY"\')::json;'
     )
     op.alter_column("transcript", "summary", new_column_name="long_summary")
     op.add_column("transcript", sa.Column("title", sa.String(), nullable=True))
@@ -34,7 +34,7 @@ def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
     op.execute(
         "UPDATE transcript SET events = "
-        'REPLACE(events, \'"event": "LONG_SUMMARY"\', \'"event": "SUMMARY"\');'
+        'REPLACE(events::text, \'"event": "LONG_SUMMARY"\', \'"event": "SUMMARY"\')::json;'
     )
     with op.batch_alter_table("transcript", schema=None) as batch_op:
         batch_op.alter_column("long_summary", nullable=True, new_column_name="summary")

server/migrations/versions/9f5c78d352d6_datetime_timezone.py

@@ -0,0 +1,121 @@
+"""datetime timezone
+
+Revision ID: 9f5c78d352d6
+Revises: 8120ebc75366
+Create Date: 2025-08-13 19:18:27.113593
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "9f5c78d352d6"
+down_revision: Union[str, None] = "8120ebc75366"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column(
+            "start_date",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+        batch_op.alter_column(
+            "end_date",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.alter_column(
+            "consent_timestamp",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.alter_column(
+            "recorded_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=postgresql.TIMESTAMP(),
+            type_=sa.DateTime(timezone=True),
+            existing_nullable=True,
+        )
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.alter_column(
+            "created_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("recording", schema=None) as batch_op:
+        batch_op.alter_column(
+            "recorded_at",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("meeting_consent", schema=None) as batch_op:
+        batch_op.alter_column(
+            "consent_timestamp",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=False,
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column(
+            "end_date",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+        batch_op.alter_column(
+            "start_date",
+            existing_type=sa.DateTime(timezone=True),
+            type_=postgresql.TIMESTAMP(),
+            existing_nullable=True,
+        )
+
+    # ### end Alembic commands ###

server/migrations/versions/a7122bc0b2ca_add_shared_rooms.py

@@ -25,7 +25,7 @@ def upgrade() -> None:
         sa.Column(
             "is_shared",
             sa.Boolean(),
-            server_default=sa.text("0"),
+            server_default=sa.text("false"),
             nullable=False,
         ),
     )

server/migrations/versions/b0e5f7876032_add_meeting_is_active.py

@@ -23,7 +23,10 @@ def upgrade() -> None:
     with op.batch_alter_table("meeting", schema=None) as batch_op:
         batch_op.add_column(
             sa.Column(
-                "is_active", sa.Boolean(), server_default=sa.text("1"), nullable=False
+                "is_active",
+                sa.Boolean(),
+                server_default=sa.text("true"),
+                nullable=False,
             )
         )
 

server/migrations/versions/b1c33bd09963_add_search_optimization_indexes.py

@@ -0,0 +1,41 @@
+"""add_search_optimization_indexes
+
+Revision ID: b1c33bd09963
+Revises: 9f5c78d352d6
+Create Date: 2025-08-14 17:26:02.117408
+
+"""
+
+from typing import Sequence, Union
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "b1c33bd09963"
+down_revision: Union[str, None] = "9f5c78d352d6"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Add indexes for actual search filtering patterns used in frontend
+    # Based on /browse page filters: room_id and source_kind
+
+    # Index for room_id + created_at (for room-specific searches with date ordering)
+    op.create_index(
+        "idx_transcript_room_id_created_at",
+        "transcript",
+        ["room_id", "created_at"],
+        if_not_exists=True,
+    )
+
+    # Index for source_kind alone (actively used filter in frontend)
+    op.create_index(
+        "idx_transcript_source_kind", "transcript", ["source_kind"], if_not_exists=True
+    )
+
+
+def downgrade() -> None:
+    # Remove the indexes in reverse order
+    op.drop_index("idx_transcript_source_kind", "transcript", if_exists=True)
+    op.drop_index("idx_transcript_room_id_created_at", "transcript", if_exists=True)

server/migrations/versions/b9348748bbbc_reviewed.py

@@ -23,7 +23,7 @@ def upgrade() -> None:
     op.add_column(
         "transcript",
         sa.Column(
-            "reviewed", sa.Boolean(), server_default=sa.text("0"), nullable=False
+            "reviewed", sa.Boolean(), server_default=sa.text("false"), nullable=False
         ),
     )
     # ### end Alembic commands ###

server/migrations/versions/d4a1c446458c_add_grace_period_fields_to_meeting.py

@@ -0,0 +1,34 @@
+"""add_grace_period_fields_to_meeting
+
+Revision ID: d4a1c446458c
+Revises: 6025e9b2bef2
+Create Date: 2025-08-18 18:50:37.768052
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "d4a1c446458c"
+down_revision: Union[str, None] = "6025e9b2bef2"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Add fields to track when participants left for grace period logic
+    op.add_column(
+        "meeting", sa.Column("last_participant_left_at", sa.DateTime(timezone=True))
+    )
+    op.add_column(
+        "meeting",
+        sa.Column("grace_period_minutes", sa.Integer, server_default=sa.text("15")),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("meeting", "grace_period_minutes")
+    op.drop_column("meeting", "last_participant_left_at")

server/migrations/versions/d8e204bbf615_add_calendar.py

@@ -0,0 +1,129 @@
+"""add calendar
+
+Revision ID: d8e204bbf615
+Revises: d4a1c446458c
+Create Date: 2025-09-10 19:56:22.295756
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+# revision identifiers, used by Alembic.
+revision: str = "d8e204bbf615"
+down_revision: Union[str, None] = "d4a1c446458c"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    op.create_table(
+        "calendar_event",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("room_id", sa.String(), nullable=False),
+        sa.Column("ics_uid", sa.Text(), nullable=False),
+        sa.Column("title", sa.Text(), nullable=True),
+        sa.Column("description", sa.Text(), nullable=True),
+        sa.Column("start_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("end_time", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("attendees", postgresql.JSONB(astext_type=sa.Text()), nullable=True),
+        sa.Column("location", sa.Text(), nullable=True),
+        sa.Column("ics_raw_data", sa.Text(), nullable=True),
+        sa.Column("last_synced", sa.DateTime(timezone=True), nullable=False),
+        sa.Column(
+            "is_deleted", sa.Boolean(), server_default=sa.text("false"), nullable=False
+        ),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("updated_at", sa.DateTime(timezone=True), nullable=False),
+        sa.ForeignKeyConstraint(
+            ["room_id"],
+            ["room.id"],
+            name="fk_calendar_event_room_id",
+            ondelete="CASCADE",
+        ),
+        sa.PrimaryKeyConstraint("id"),
+        sa.UniqueConstraint("room_id", "ics_uid", name="uq_room_calendar_event"),
+    )
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.create_index(
+            "idx_calendar_event_deleted",
+            ["is_deleted"],
+            unique=False,
+            postgresql_where=sa.text("NOT is_deleted"),
+        )
+        batch_op.create_index(
+            "idx_calendar_event_room_start", ["room_id", "start_time"], unique=False
+        )
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("calendar_event_id", sa.String(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "calendar_metadata",
+                postgresql.JSONB(astext_type=sa.Text()),
+                nullable=True,
+            )
+        )
+        batch_op.create_index(
+            "idx_meeting_calendar_event", ["calendar_event_id"], unique=False
+        )
+        batch_op.create_foreign_key(
+            "fk_meeting_calendar_event_id",
+            "calendar_event",
+            ["calendar_event_id"],
+            ["id"],
+            ondelete="SET NULL",
+        )
+
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.add_column(sa.Column("ics_url", sa.Text(), nullable=True))
+        batch_op.add_column(
+            sa.Column(
+                "ics_fetch_interval", sa.Integer(), server_default="300", nullable=True
+            )
+        )
+        batch_op.add_column(
+            sa.Column(
+                "ics_enabled",
+                sa.Boolean(),
+                server_default=sa.text("false"),
+                nullable=False,
+            )
+        )
+        batch_op.add_column(
+            sa.Column("ics_last_sync", sa.DateTime(timezone=True), nullable=True)
+        )
+        batch_op.add_column(sa.Column("ics_last_etag", sa.Text(), nullable=True))
+        batch_op.create_index("idx_room_ics_enabled", ["ics_enabled"], unique=False)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("room", schema=None) as batch_op:
+        batch_op.drop_index("idx_room_ics_enabled")
+        batch_op.drop_column("ics_last_etag")
+        batch_op.drop_column("ics_last_sync")
+        batch_op.drop_column("ics_enabled")
+        batch_op.drop_column("ics_fetch_interval")
+        batch_op.drop_column("ics_url")
+
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_constraint("fk_meeting_calendar_event_id", type_="foreignkey")
+        batch_op.drop_index("idx_meeting_calendar_event")
+        batch_op.drop_column("calendar_metadata")
+        batch_op.drop_column("calendar_event_id")
+
+    with op.batch_alter_table("calendar_event", schema=None) as batch_op:
+        batch_op.drop_index("idx_calendar_event_room_start")
+        batch_op.drop_index(
+            "idx_calendar_event_deleted", postgresql_where=sa.text("NOT is_deleted")
+        )
+
+    op.drop_table("calendar_event")
+    # ### end Alembic commands ###

server/migrations/versions/dc035ff72fd5_remove_grace_period_fields.py

@@ -0,0 +1,43 @@
+"""remove_grace_period_fields
+
+Revision ID: dc035ff72fd5
+Revises: d8e204bbf615
+Create Date: 2025-09-11 10:36:45.197588
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "dc035ff72fd5"
+down_revision: Union[str, None] = "d8e204bbf615"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # Remove grace period columns from meeting table
+    op.drop_column("meeting", "last_participant_left_at")
+    op.drop_column("meeting", "grace_period_minutes")
+
+
+def downgrade() -> None:
+    # Add back grace period columns to meeting table
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "last_participant_left_at", sa.DateTime(timezone=True), nullable=True
+        ),
+    )
+    op.add_column(
+        "meeting",
+        sa.Column(
+            "grace_period_minutes",
+            sa.Integer(),
+            server_default=sa.text("15"),
+            nullable=True,
+        ),
+    )

server/migrations/versions/def1b5867d4c_make_meeting_room_id_nullable_but_keep_.py

@@ -0,0 +1,34 @@
+"""make meeting room_id nullable but keep foreign key
+
+Revision ID: def1b5867d4c
+Revises: 0ce521cda2ee
+Create Date: 2025-09-11 09:42:18.697264
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "def1b5867d4c"
+down_revision: Union[str, None] = "0ce521cda2ee"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=True)
+
+    # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+    # ### commands auto generated by Alembic - please adjust! ###
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.alter_column("room_id", existing_type=sa.VARCHAR(), nullable=False)
+
+    # ### end Alembic commands ###

server/pyproject.toml

@@ -12,7 +12,6 @@ dependencies = [
     "requests>=2.31.0",
     "aiortc>=1.5.0",
     "sortedcontainers>=2.4.0",
-    "loguru>=0.7.0",
     "pydantic-settings>=2.0.2",
     "structlog>=23.1.0",
     "uvicorn[standard]>=0.23.1",
@@ -20,26 +19,26 @@ dependencies = [
     "sentry-sdk[fastapi]>=1.29.2",
     "httpx>=0.24.1",
     "fastapi-pagination>=0.12.6",
-    "databases[aiosqlite, asyncpg]>=0.7.0",
-    "sqlalchemy<1.5",
+    "sqlalchemy>=2.0.0",
+    "asyncpg>=0.29.0",
     "alembic>=1.11.3",
     "nltk>=3.8.1",
     "prometheus-fastapi-instrumentator>=6.1.0",
     "sentencepiece>=0.1.99",
     "protobuf>=4.24.3",
-    "profanityfilter>=2.0.6",
     "celery>=5.3.4",
     "redis>=5.0.1",
     "python-jose[cryptography]>=3.3.0",
     "python-multipart>=0.0.6",
-    "faster-whisper>=0.10.0",
     "transformers>=4.36.2",
-    "black==24.1.1",
     "jsonschema>=4.23.0",
     "openai>=1.59.7",
     "psycopg2-binary>=2.9.10",
     "llama-index>=0.12.52",
     "llama-index-llms-openai-like>=0.4.0",
+    "pytest-env>=1.1.5",
+    "webvtt-py>=0.5.0",
+    "icalendar>=6.0.0",
 ]
 
 [dependency-groups]
@@ -47,6 +46,7 @@ dev = [
     "black>=24.1.1",
     "stamina>=23.1.0",
     "pyinstrument>=4.6.1",
+    "pytest-async-sqlalchemy>=0.2.0",
 ]
 tests = [
     "pytest-cov>=4.1.0",
@@ -56,6 +56,9 @@ tests = [
     "httpx-ws>=0.4.1",
     "pytest-httpx>=0.23.1",
     "pytest-celery>=0.0.0",
+    "pytest-recording>=0.13.4",
+    "pytest-docker>=3.2.3",
+    "asgi-lifespan>=2.1.0",
 ]
 aws = ["aioboto3>=11.2.0"]
 evaluation = [
@@ -64,6 +67,15 @@ evaluation = [
     "tqdm>=4.66.0",
     "pydantic>=2.1.1",
 ]
+local = [
+    "pyannote-audio>=3.3.2",
+    "faster-whisper>=0.10.0",
+]
+silero-vad = [
+    "silero-vad>=5.1.2",
+    "torch>=2.8.0",
+    "torchaudio>=2.8.0",
+]
 
 [tool.uv]
 default-groups = [
@@ -71,6 +83,21 @@ default-groups = [
     "tests",
     "aws",
     "evaluation",
+    "local",
+    "silero-vad"
+]
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[tool.uv.sources]
+torch = [
+  { index = "pytorch-cpu" },
+]
+torchaudio = [
+  { index = "pytorch-cpu" },
 ]
 
 [build-system]
@@ -83,10 +110,31 @@ packages = ["reflector"]
 [tool.coverage.run]
 source = ["reflector"]
 
+[tool.pytest_env]
+ENVIRONMENT = "pytest"
+DATABASE_URL = "postgresql+asyncpg://test_user:test_password@localhost:15432/reflector_test"
+
 [tool.pytest.ini_options]
 addopts = "-ra -q --disable-pytest-warnings --cov --cov-report html -v"
 testpaths = ["tests"]
 asyncio_mode = "auto"
+asyncio_debug = true
+asyncio_default_fixture_loop_scope = "session"
+asyncio_default_test_loop_scope = "session"
+markers = [
+    "model_api: tests for the unified model-serving HTTP API (backend- and hardware-agnostic)",
+]
+
+[tool.ruff.lint]
+select = [
+    "I",       # isort - import sorting
+    "F401",    # unused imports
+    "PLC0415", # import-outside-top-level - detect inline imports
+]
 
 [tool.ruff.lint.per-file-ignores]
 "reflector/processors/summary/summary_builder.py" = ["E501"]
+"gpu/modal_deployments/**.py" = ["PLC0415"]
+"reflector/tools/**.py" = ["PLC0415"]
+"migrations/versions/**.py" = ["PLC0415"]
+"tests/**.py" = ["PLC0415"]

server/reflector/asynctask.py

@@ -0,0 +1,20 @@
+import asyncio
+import functools
+
+
+def asynctask(f):
+    @functools.wraps(f)
+    def wrapper(*args, **kwargs):
+        async def run_async():
+            return await f(*args, **kwargs)
+
+        coro = run_async()
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            loop = None
+        if loop and loop.is_running():
+            return loop.run_until_complete(coro)
+        return asyncio.run(coro)
+
+    return wrapper

server/reflector/auth/auth_jwt.py

@@ -67,7 +67,8 @@ def current_user(
     try:
         payload = jwtauth.verify_token(token)
         sub = payload["sub"]
-        return UserInfo(sub=sub)
+        email = payload["email"]
+        return UserInfo(sub=sub, email=email)
     except JWTError as e:
         logger.error(f"JWT error: {e}")
         raise HTTPException(status_code=401, detail="Invalid authentication")

server/reflector/db/__init__.py

@@ -1,29 +1,69 @@
-import databases
-import sqlalchemy
+from typing import AsyncGenerator
 
+from sqlalchemy.ext.asyncio import (
+    AsyncEngine,
+    AsyncSession,
+    async_sessionmaker,
+    create_async_engine,
+)
+
+from reflector.db.base import Base as Base
+from reflector.db.base import metadata as metadata
 from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.settings import settings
 
-database = databases.Database(settings.DATABASE_URL)
-metadata = sqlalchemy.MetaData()
+_engine: AsyncEngine | None = None
+_session_factory: async_sessionmaker[AsyncSession] | None = None
 
-# import models
+
+def get_engine() -> AsyncEngine:
+    global _engine
+    if _engine is None:
+        _engine = create_async_engine(
+            settings.DATABASE_URL,
+            echo=False,
+            pool_pre_ping=True,
+        )
+    return _engine
+
+
+def get_session_factory() -> async_sessionmaker[AsyncSession]:
+    global _session_factory
+    if _session_factory is None:
+        _session_factory = async_sessionmaker(
+            get_engine(),
+            class_=AsyncSession,
+            expire_on_commit=False,
+        )
+    return _session_factory
+
+
+async def _get_session() -> AsyncGenerator[AsyncSession, None]:
+    # necessary implementation to ease mocking on pytest
+    async with get_session_factory()() as session:
+        yield session
+
+
+async def get_session() -> AsyncGenerator[AsyncSession, None]:
+    async for session in _get_session():
+        yield session
+
+
+import reflector.db.calendar_events  # noqa
 import reflector.db.meetings  # noqa
 import reflector.db.recordings  # noqa
 import reflector.db.rooms  # noqa
 import reflector.db.transcripts  # noqa
 
-kwargs = {}
-if "sqlite" in settings.DATABASE_URL:
-    kwargs["connect_args"] = {"check_same_thread": False}
-engine = sqlalchemy.create_engine(settings.DATABASE_URL, **kwargs)
-
 
 @subscribers_startup.append
 async def database_connect(_):
-    await database.connect()
+    get_engine()
 
 
 @subscribers_shutdown.append
 async def database_disconnect(_):
-    await database.disconnect()
+    global _engine
+    if _engine:
+        await _engine.dispose()
+        _engine = None

server/reflector/db/base.py

@@ -0,0 +1,237 @@
+from datetime import datetime
+from typing import Optional
+
+import sqlalchemy as sa
+from sqlalchemy.dialects.postgresql import JSONB, TSVECTOR
+from sqlalchemy.ext.asyncio import AsyncAttrs
+from sqlalchemy.orm import DeclarativeBase, Mapped, mapped_column
+
+
+class Base(AsyncAttrs, DeclarativeBase):
+    pass
+
+
+class TranscriptModel(Base):
+    __tablename__ = "transcript"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    name: Mapped[Optional[str]] = mapped_column(sa.String)
+    status: Mapped[Optional[str]] = mapped_column(sa.String)
+    locked: Mapped[Optional[bool]] = mapped_column(sa.Boolean)
+    duration: Mapped[Optional[float]] = mapped_column(sa.Float)
+    created_at: Mapped[Optional[datetime]] = mapped_column(sa.DateTime(timezone=True))
+    title: Mapped[Optional[str]] = mapped_column(sa.String)
+    short_summary: Mapped[Optional[str]] = mapped_column(sa.String)
+    long_summary: Mapped[Optional[str]] = mapped_column(sa.String)
+    topics: Mapped[Optional[list]] = mapped_column(sa.JSON)
+    events: Mapped[Optional[list]] = mapped_column(sa.JSON)
+    participants: Mapped[Optional[list]] = mapped_column(sa.JSON)
+    source_language: Mapped[Optional[str]] = mapped_column(sa.String)
+    target_language: Mapped[Optional[str]] = mapped_column(sa.String)
+    reviewed: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    audio_location: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="local"
+    )
+    user_id: Mapped[Optional[str]] = mapped_column(sa.String)
+    share_mode: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="private"
+    )
+    meeting_id: Mapped[Optional[str]] = mapped_column(sa.String)
+    recording_id: Mapped[Optional[str]] = mapped_column(sa.String)
+    zulip_message_id: Mapped[Optional[int]] = mapped_column(sa.Integer)
+    source_kind: Mapped[str] = mapped_column(
+        sa.String, nullable=False
+    )  # Enum will be handled separately
+    audio_deleted: Mapped[Optional[bool]] = mapped_column(sa.Boolean)
+    room_id: Mapped[Optional[str]] = mapped_column(sa.String)
+    webvtt: Mapped[Optional[str]] = mapped_column(sa.Text)
+
+    __table_args__ = (
+        sa.Index("idx_transcript_recording_id", "recording_id"),
+        sa.Index("idx_transcript_user_id", "user_id"),
+        sa.Index("idx_transcript_created_at", "created_at"),
+        sa.Index("idx_transcript_user_id_recording_id", "user_id", "recording_id"),
+        sa.Index("idx_transcript_room_id", "room_id"),
+        sa.Index("idx_transcript_source_kind", "source_kind"),
+        sa.Index("idx_transcript_room_id_created_at", "room_id", "created_at"),
+    )
+
+
+TranscriptModel.search_vector_en = sa.Column(
+    "search_vector_en",
+    TSVECTOR,
+    sa.Computed(
+        "setweight(to_tsvector('english', coalesce(title, '')), 'A') || "
+        "setweight(to_tsvector('english', coalesce(long_summary, '')), 'B') || "
+        "setweight(to_tsvector('english', coalesce(webvtt, '')), 'C')",
+        persisted=True,
+    ),
+)
+
+
+class RoomModel(Base):
+    __tablename__ = "room"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    name: Mapped[str] = mapped_column(sa.String, nullable=False, unique=True)
+    user_id: Mapped[str] = mapped_column(sa.String, nullable=False)
+    created_at: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+    zulip_auto_post: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    zulip_stream: Mapped[Optional[str]] = mapped_column(sa.String)
+    zulip_topic: Mapped[Optional[str]] = mapped_column(sa.String)
+    is_locked: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    room_mode: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="normal"
+    )
+    recording_type: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="cloud"
+    )
+    recording_trigger: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="automatic-2nd-participant"
+    )
+    is_shared: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    webhook_url: Mapped[Optional[str]] = mapped_column(sa.String)
+    webhook_secret: Mapped[Optional[str]] = mapped_column(sa.String)
+    ics_url: Mapped[Optional[str]] = mapped_column(sa.Text)
+    ics_fetch_interval: Mapped[Optional[int]] = mapped_column(
+        sa.Integer, server_default=sa.text("300")
+    )
+    ics_enabled: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    ics_last_sync: Mapped[Optional[datetime]] = mapped_column(
+        sa.DateTime(timezone=True)
+    )
+    ics_last_etag: Mapped[Optional[str]] = mapped_column(sa.Text)
+
+    __table_args__ = (
+        sa.Index("idx_room_is_shared", "is_shared"),
+        sa.Index("idx_room_ics_enabled", "ics_enabled"),
+    )
+
+
+class MeetingModel(Base):
+    __tablename__ = "meeting"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    room_name: Mapped[Optional[str]] = mapped_column(sa.String)
+    room_url: Mapped[Optional[str]] = mapped_column(sa.String)
+    host_room_url: Mapped[Optional[str]] = mapped_column(sa.String)
+    start_date: Mapped[Optional[datetime]] = mapped_column(sa.DateTime(timezone=True))
+    end_date: Mapped[Optional[datetime]] = mapped_column(sa.DateTime(timezone=True))
+    room_id: Mapped[Optional[str]] = mapped_column(
+        sa.String, sa.ForeignKey("room.id", ondelete="CASCADE")
+    )
+    is_locked: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    room_mode: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="normal"
+    )
+    recording_type: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="cloud"
+    )
+    recording_trigger: Mapped[str] = mapped_column(
+        sa.String, nullable=False, server_default="automatic-2nd-participant"
+    )
+    num_clients: Mapped[int] = mapped_column(
+        sa.Integer, nullable=False, server_default=sa.text("0")
+    )
+    is_active: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("true")
+    )
+    calendar_event_id: Mapped[Optional[str]] = mapped_column(
+        sa.String,
+        sa.ForeignKey(
+            "calendar_event.id",
+            ondelete="SET NULL",
+            name="fk_meeting_calendar_event_id",
+        ),
+    )
+    calendar_metadata: Mapped[Optional[dict]] = mapped_column(JSONB)
+
+    __table_args__ = (
+        sa.Index("idx_meeting_room_id", "room_id"),
+        sa.Index("idx_meeting_calendar_event", "calendar_event_id"),
+    )
+
+
+class MeetingConsentModel(Base):
+    __tablename__ = "meeting_consent"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    meeting_id: Mapped[str] = mapped_column(
+        sa.String, sa.ForeignKey("meeting.id", ondelete="CASCADE"), nullable=False
+    )
+    user_id: Mapped[Optional[str]] = mapped_column(sa.String)
+    consent_given: Mapped[bool] = mapped_column(sa.Boolean, nullable=False)
+    consent_timestamp: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+
+
+class RecordingModel(Base):
+    __tablename__ = "recording"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    meeting_id: Mapped[str] = mapped_column(
+        sa.String, sa.ForeignKey("meeting.id", ondelete="CASCADE"), nullable=False
+    )
+    url: Mapped[str] = mapped_column(sa.String, nullable=False)
+    object_key: Mapped[str] = mapped_column(sa.String, nullable=False)
+    duration: Mapped[Optional[float]] = mapped_column(sa.Float)
+    created_at: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+
+    __table_args__ = (sa.Index("idx_recording_meeting_id", "meeting_id"),)
+
+
+class CalendarEventModel(Base):
+    __tablename__ = "calendar_event"
+
+    id: Mapped[str] = mapped_column(sa.String, primary_key=True)
+    room_id: Mapped[str] = mapped_column(
+        sa.String, sa.ForeignKey("room.id", ondelete="CASCADE"), nullable=False
+    )
+    ics_uid: Mapped[str] = mapped_column(sa.Text, nullable=False)
+    title: Mapped[Optional[str]] = mapped_column(sa.Text)
+    description: Mapped[Optional[str]] = mapped_column(sa.Text)
+    start_time: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+    end_time: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+    attendees: Mapped[Optional[dict]] = mapped_column(JSONB)
+    location: Mapped[Optional[str]] = mapped_column(sa.Text)
+    ics_raw_data: Mapped[Optional[str]] = mapped_column(sa.Text)
+    last_synced: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+    is_deleted: Mapped[bool] = mapped_column(
+        sa.Boolean, nullable=False, server_default=sa.text("false")
+    )
+    created_at: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+    updated_at: Mapped[datetime] = mapped_column(
+        sa.DateTime(timezone=True), nullable=False
+    )
+
+    __table_args__ = (
+        sa.Index("idx_calendar_event_room_start", "room_id", "start_time"),
+    )
+
+
+metadata = Base.metadata

server/reflector/db/calendar_events.py

@@ -0,0 +1,187 @@
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+import sqlalchemy as sa
+from pydantic import BaseModel, ConfigDict, Field
+from sqlalchemy import delete, select, update
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from reflector.db.base import CalendarEventModel
+from reflector.utils import generate_uuid4
+
+
+class CalendarEvent(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
+    id: str = Field(default_factory=generate_uuid4)
+    room_id: str
+    ics_uid: str
+    title: str | None = None
+    description: str | None = None
+    start_time: datetime
+    end_time: datetime
+    attendees: list[dict[str, Any]] | None = None
+    location: str | None = None
+    ics_raw_data: str | None = None
+    last_synced: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    is_deleted: bool = False
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+    updated_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
+
+
+class CalendarEventController:
+    async def get_upcoming_events(
+        self,
+        session: AsyncSession,
+        room_id: str,
+        current_time: datetime,
+        buffer_minutes: int = 15,
+    ) -> list[CalendarEvent]:
+        buffer_time = current_time + timedelta(minutes=buffer_minutes)
+
+        query = (
+            select(CalendarEventModel)
+            .where(
+                sa.and_(
+                    CalendarEventModel.room_id == room_id,
+                    CalendarEventModel.start_time <= buffer_time,
+                    CalendarEventModel.end_time > current_time,
+                )
+            )
+            .order_by(CalendarEventModel.start_time)
+        )
+
+        result = await session.execute(query)
+        return [CalendarEvent.model_validate(row) for row in result.scalars().all()]
+
+    async def get_by_id(
+        self, session: AsyncSession, event_id: str
+    ) -> CalendarEvent | None:
+        query = select(CalendarEventModel).where(CalendarEventModel.id == event_id)
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
+            return None
+        return CalendarEvent.model_validate(row)
+
+    async def get_by_ics_uid(
+        self, session: AsyncSession, room_id: str, ics_uid: str
+    ) -> CalendarEvent | None:
+        query = select(CalendarEventModel).where(
+            sa.and_(
+                CalendarEventModel.room_id == room_id,
+                CalendarEventModel.ics_uid == ics_uid,
+            )
+        )
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
+            return None
+        return CalendarEvent.model_validate(row)
+
+    async def upsert(
+        self, session: AsyncSession, event: CalendarEvent
+    ) -> CalendarEvent:
+        existing = await self.get_by_ics_uid(session, event.room_id, event.ics_uid)
+
+        if existing:
+            event.updated_at = datetime.now(timezone.utc)
+            query = (
+                update(CalendarEventModel)
+                .where(CalendarEventModel.id == existing.id)
+                .values(**event.model_dump(exclude={"id"}))
+            )
+            await session.execute(query)
+            await session.commit()
+            return event
+        else:
+            new_event = CalendarEventModel(**event.model_dump())
+            session.add(new_event)
+            await session.commit()
+            return event
+
+    async def delete_old_events(
+        self, session: AsyncSession, room_id: str, cutoff_date: datetime
+    ) -> int:
+        query = delete(CalendarEventModel).where(
+            sa.and_(
+                CalendarEventModel.room_id == room_id,
+                CalendarEventModel.end_time < cutoff_date,
+            )
+        )
+        result = await session.execute(query)
+        await session.commit()
+        return result.rowcount
+
+    async def delete_events_not_in_list(
+        self, session: AsyncSession, room_id: str, keep_ics_uids: list[str]
+    ) -> int:
+        if not keep_ics_uids:
+            query = delete(CalendarEventModel).where(
+                CalendarEventModel.room_id == room_id
+            )
+        else:
+            query = delete(CalendarEventModel).where(
+                sa.and_(
+                    CalendarEventModel.room_id == room_id,
+                    CalendarEventModel.ics_uid.notin_(keep_ics_uids),
+                )
+            )
+
+        result = await session.execute(query)
+        await session.commit()
+        return result.rowcount
+
+    async def get_by_room(
+        self, session: AsyncSession, room_id: str, include_deleted: bool = True
+    ) -> list[CalendarEvent]:
+        query = select(CalendarEventModel).where(CalendarEventModel.room_id == room_id)
+        if not include_deleted:
+            query = query.where(CalendarEventModel.is_deleted == False)
+        result = await session.execute(query)
+        return [CalendarEvent.model_validate(row) for row in result.scalars().all()]
+
+    async def get_upcoming(
+        self, session: AsyncSession, room_id: str, minutes_ahead: int = 120
+    ) -> list[CalendarEvent]:
+        now = datetime.now(timezone.utc)
+        buffer_time = now + timedelta(minutes=minutes_ahead)
+
+        query = (
+            select(CalendarEventModel)
+            .where(
+                sa.and_(
+                    CalendarEventModel.room_id == room_id,
+                    CalendarEventModel.start_time <= buffer_time,
+                    CalendarEventModel.end_time > now,
+                    CalendarEventModel.is_deleted == False,
+                )
+            )
+            .order_by(CalendarEventModel.start_time)
+        )
+
+        result = await session.execute(query)
+        return [CalendarEvent.model_validate(row) for row in result.scalars().all()]
+
+    async def soft_delete_missing(
+        self, session: AsyncSession, room_id: str, current_ics_uids: list[str]
+    ) -> int:
+        query = (
+            update(CalendarEventModel)
+            .where(
+                sa.and_(
+                    CalendarEventModel.room_id == room_id,
+                    CalendarEventModel.ics_uid.notin_(current_ics_uids)
+                    if current_ics_uids
+                    else True,
+                    CalendarEventModel.end_time > datetime.now(timezone.utc),
+                )
+            )
+            .values(is_deleted=True)
+        )
+        result = await session.execute(query)
+        await session.commit()
+        return result.rowcount
+
+
+calendar_events_controller = CalendarEventController()

server/reflector/db/meetings.py

@@ -1,61 +1,19 @@
 from datetime import datetime
-from typing import Literal
+from typing import Any, Literal
 
 import sqlalchemy as sa
-from fastapi import HTTPException
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
+from sqlalchemy import select, update
+from sqlalchemy.ext.asyncio import AsyncSession
 
-from reflector.db import database, metadata
+from reflector.db.base import MeetingConsentModel, MeetingModel
 from reflector.db.rooms import Room
 from reflector.utils import generate_uuid4
 
-meetings = sa.Table(
-    "meeting",
-    metadata,
-    sa.Column("id", sa.String, primary_key=True),
-    sa.Column("room_name", sa.String),
-    sa.Column("room_url", sa.String),
-    sa.Column("host_room_url", sa.String),
-    sa.Column("start_date", sa.DateTime),
-    sa.Column("end_date", sa.DateTime),
-    sa.Column("user_id", sa.String),
-    sa.Column("room_id", sa.String),
-    sa.Column("is_locked", sa.Boolean, nullable=False, server_default=sa.false()),
-    sa.Column("room_mode", sa.String, nullable=False, server_default="normal"),
-    sa.Column("recording_type", sa.String, nullable=False, server_default="cloud"),
-    sa.Column(
-        "recording_trigger",
-        sa.String,
-        nullable=False,
-        server_default="automatic-2nd-participant",
-    ),
-    sa.Column(
-        "num_clients",
-        sa.Integer,
-        nullable=False,
-        server_default=sa.text("0"),
-    ),
-    sa.Column(
-        "is_active",
-        sa.Boolean,
-        nullable=False,
-        server_default=sa.true(),
-    ),
-    sa.Index("idx_meeting_room_id", "room_id"),
-)
-
-meeting_consent = sa.Table(
-    "meeting_consent",
-    metadata,
-    sa.Column("id", sa.String, primary_key=True),
-    sa.Column("meeting_id", sa.String, sa.ForeignKey("meeting.id"), nullable=False),
-    sa.Column("user_id", sa.String),
-    sa.Column("consent_given", sa.Boolean, nullable=False),
-    sa.Column("consent_timestamp", sa.DateTime, nullable=False),
-)
-
 
 class MeetingConsent(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
     id: str = Field(default_factory=generate_uuid4)
     meeting_id: str
     user_id: str | None = None
@@ -64,14 +22,15 @@ class MeetingConsent(BaseModel):
 
 
 class Meeting(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
     id: str
     room_name: str
     room_url: str
     host_room_url: str
     start_date: datetime
     end_date: datetime
-    user_id: str | None = None
-    room_id: str | None = None
+    room_id: str | None
     is_locked: bool = False
     room_mode: Literal["normal", "group"] = "normal"
     recording_type: Literal["none", "local", "cloud"] = "cloud"
@@ -79,23 +38,25 @@ class Meeting(BaseModel):
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     num_clients: int = 0
+    is_active: bool = True
+    calendar_event_id: str | None = None
+    calendar_metadata: dict[str, Any] | None = None
 
 
 class MeetingController:
     async def create(
         self,
+        session: AsyncSession,
         id: str,
         room_name: str,
         room_url: str,
         host_room_url: str,
         start_date: datetime,
         end_date: datetime,
-        user_id: str,
         room: Room,
+        calendar_event_id: str | None = None,
+        calendar_metadata: dict[str, Any] | None = None,
     ):
-        """
-        Create a new meeting
-        """
         meeting = Meeting(
             id=id,
             room_name=room_name,
@@ -103,148 +64,206 @@ class MeetingController:
             host_room_url=host_room_url,
             start_date=start_date,
             end_date=end_date,
-            user_id=user_id,
             room_id=room.id,
             is_locked=room.is_locked,
             room_mode=room.room_mode,
             recording_type=room.recording_type,
             recording_trigger=room.recording_trigger,
+            calendar_event_id=calendar_event_id,
+            calendar_metadata=calendar_metadata,
         )
-        query = meetings.insert().values(**meeting.model_dump())
-        await database.execute(query)
+        new_meeting = MeetingModel(**meeting.model_dump())
+        session.add(new_meeting)
+        await session.commit()
         return meeting
 
-    async def get_all_active(self) -> list[Meeting]:
-        """
-        Get active meetings.
-        """
-        query = meetings.select().where(meetings.c.is_active)
-        return await database.fetch_all(query)
+    async def get_all_active(self, session: AsyncSession) -> list[Meeting]:
+        query = select(MeetingModel).where(MeetingModel.is_active)
+        result = await session.execute(query)
+        return [Meeting.model_validate(row) for row in result.scalars().all()]
 
     async def get_by_room_name(
         self,
+        session: AsyncSession,
         room_name: str,
-    ) -> Meeting:
+    ) -> Meeting | None:
         """
         Get a meeting by room name.
+        For backward compatibility, returns the most recent meeting.
         """
-        query = meetings.select().where(meetings.c.room_name == room_name)
-        result = await database.fetch_one(query)
-        if not result:
+        query = (
+            select(MeetingModel)
+            .where(MeetingModel.room_name == room_name)
+            .order_by(MeetingModel.end_date.desc())
+        )
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             return None
+        return Meeting.model_validate(row)
 
-        return Meeting(**result)
-
-    async def get_active(self, room: Room, current_time: datetime) -> Meeting:
+    async def get_active(
+        self, session: AsyncSession, room: Room, current_time: datetime
+    ) -> Meeting | None:
         """
         Get latest active meeting for a room.
+        For backward compatibility, returns the most recent active meeting.
         """
-        end_date = getattr(meetings.c, "end_date")
         query = (
-            meetings.select()
+            select(MeetingModel)
             .where(
                 sa.and_(
-                    meetings.c.room_id == room.id,
-                    meetings.c.end_date > current_time,
-                    meetings.c.is_active,
+                    MeetingModel.room_id == room.id,
+                    MeetingModel.end_date > current_time,
+                    MeetingModel.is_active,
                 )
             )
-            .order_by(end_date.desc())
+            .order_by(MeetingModel.end_date.desc())
         )
-        result = await database.fetch_one(query)
-        if not result:
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             return None
+        return Meeting.model_validate(row)
 
-        return Meeting(**result)
+    async def get_all_active_for_room(
+        self, session: AsyncSession, room: Room, current_time: datetime
+    ) -> list[Meeting]:
+        query = (
+            select(MeetingModel)
+            .where(
+                sa.and_(
+                    MeetingModel.room_id == room.id,
+                    MeetingModel.end_date > current_time,
+                    MeetingModel.is_active,
+                )
+            )
+            .order_by(MeetingModel.end_date.desc())
+        )
+        result = await session.execute(query)
+        return [Meeting.model_validate(row) for row in result.scalars().all()]
 
-    async def get_by_id(self, meeting_id: str, **kwargs) -> Meeting | None:
+    async def get_active_by_calendar_event(
+        self,
+        session: AsyncSession,
+        room: Room,
+        calendar_event_id: str,
+        current_time: datetime,
+    ) -> Meeting | None:
         """
-        Get a meeting by id
+        Get active meeting for a specific calendar event.
         """
-        query = meetings.select().where(meetings.c.id == meeting_id)
-        result = await database.fetch_one(query)
-        if not result:
+        query = select(MeetingModel).where(
+            sa.and_(
+                MeetingModel.room_id == room.id,
+                MeetingModel.calendar_event_id == calendar_event_id,
+                MeetingModel.end_date > current_time,
+                MeetingModel.is_active,
+            )
+        )
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             return None
-        return Meeting(**result)
+        return Meeting.model_validate(row)
 
-    async def get_by_id_for_http(self, meeting_id: str, user_id: str | None) -> Meeting:
-        """
-        Get a meeting by ID for HTTP request.
+    async def get_by_id(
+        self, session: AsyncSession, meeting_id: str, **kwargs
+    ) -> Meeting | None:
+        query = select(MeetingModel).where(MeetingModel.id == meeting_id)
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
+            return None
+        return Meeting.model_validate(row)
 
-        If not found, it will raise a 404 error.
-        """
-        query = meetings.select().where(meetings.c.id == meeting_id)
-        result = await database.fetch_one(query)
-        if not result:
-            raise HTTPException(status_code=404, detail="Meeting not found")
+    async def get_by_calendar_event(
+        self, session: AsyncSession, calendar_event_id: str
+    ) -> Meeting | None:
+        query = select(MeetingModel).where(
+            MeetingModel.calendar_event_id == calendar_event_id
+        )
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
+            return None
+        return Meeting.model_validate(row)
 
-        meeting = Meeting(**result)
-        if result["user_id"] != user_id:
-            meeting.host_room_url = ""
-
-        return meeting
-
-    async def update_meeting(self, meeting_id: str, **kwargs):
-        query = meetings.update().where(meetings.c.id == meeting_id).values(**kwargs)
-        await database.execute(query)
+    async def update_meeting(self, session: AsyncSession, meeting_id: str, **kwargs):
+        query = (
+            update(MeetingModel).where(MeetingModel.id == meeting_id).values(**kwargs)
+        )
+        await session.execute(query)
+        await session.commit()
 
 
 class MeetingConsentController:
-    async def get_by_meeting_id(self, meeting_id: str) -> list[MeetingConsent]:
-        query = meeting_consent.select().where(
-            meeting_consent.c.meeting_id == meeting_id
+    async def get_by_meeting_id(
+        self, session: AsyncSession, meeting_id: str
+    ) -> list[MeetingConsent]:
+        query = select(MeetingConsentModel).where(
+            MeetingConsentModel.meeting_id == meeting_id
         )
-        results = await database.fetch_all(query)
-        return [MeetingConsent(**result) for result in results]
+        result = await session.execute(query)
+        return [MeetingConsent.model_validate(row) for row in result.scalars().all()]
 
     async def get_by_meeting_and_user(
-        self, meeting_id: str, user_id: str
+        self, session: AsyncSession, meeting_id: str, user_id: str
     ) -> MeetingConsent | None:
         """Get existing consent for a specific user and meeting"""
-        query = meeting_consent.select().where(
-            meeting_consent.c.meeting_id == meeting_id,
-            meeting_consent.c.user_id == user_id,
+        query = select(MeetingConsentModel).where(
+            sa.and_(
+                MeetingConsentModel.meeting_id == meeting_id,
+                MeetingConsentModel.user_id == user_id,
+            )
         )
-        result = await database.fetch_one(query)
-        if result is None:
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if row is None:
             return None
-        return MeetingConsent(**result) if result else None
+        return MeetingConsent.model_validate(row)
 
-    async def upsert(self, consent: MeetingConsent) -> MeetingConsent:
-        """Create new consent or update existing one for authenticated users"""
+    async def upsert(
+        self, session: AsyncSession, consent: MeetingConsent
+    ) -> MeetingConsent:
         if consent.user_id:
             # For authenticated users, check if consent already exists
             # not transactional but we're ok with that; the consents ain't deleted anyways
             existing = await self.get_by_meeting_and_user(
-                consent.meeting_id, consent.user_id
+                session, consent.meeting_id, consent.user_id
             )
             if existing:
                 query = (
-                    meeting_consent.update()
-                    .where(meeting_consent.c.id == existing.id)
+                    update(MeetingConsentModel)
+                    .where(MeetingConsentModel.id == existing.id)
                     .values(
                         consent_given=consent.consent_given,
                         consent_timestamp=consent.consent_timestamp,
                     )
                 )
-                await database.execute(query)
+            await session.execute(query)
+            await session.commit()
 
-                existing.consent_given = consent.consent_given
-                existing.consent_timestamp = consent.consent_timestamp
-                return existing
+            existing.consent_given = consent.consent_given
+            existing.consent_timestamp = consent.consent_timestamp
+            return existing
 
-        query = meeting_consent.insert().values(**consent.model_dump())
-        await database.execute(query)
+        new_consent = MeetingConsentModel(**consent.model_dump())
+        session.add(new_consent)
+        await session.commit()
         return consent
 
-    async def has_any_denial(self, meeting_id: str) -> bool:
+    async def has_any_denial(self, session: AsyncSession, meeting_id: str) -> bool:
         """Check if any participant denied consent for this meeting"""
-        query = meeting_consent.select().where(
-            meeting_consent.c.meeting_id == meeting_id,
-            meeting_consent.c.consent_given.is_(False),
+        query = select(MeetingConsentModel).where(
+            sa.and_(
+                MeetingConsentModel.meeting_id == meeting_id,
+                MeetingConsentModel.consent_given.is_(False),
+            )
         )
-        result = await database.fetch_one(query)
-        return result is not None
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        return row is not None
 
 
 meetings_controller = MeetingController()

server/reflector/db/recordings.py

@@ -1,57 +1,79 @@
-from datetime import datetime
-from typing import Literal
+from datetime import datetime, timezone
 
-import sqlalchemy as sa
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, ConfigDict, Field
+from sqlalchemy import delete, select
+from sqlalchemy.ext.asyncio import AsyncSession
 
-from reflector.db import database, metadata
+from reflector.db.base import RecordingModel
 from reflector.utils import generate_uuid4
 
-recordings = sa.Table(
-    "recording",
-    metadata,
-    sa.Column("id", sa.String, primary_key=True),
-    sa.Column("bucket_name", sa.String, nullable=False),
-    sa.Column("object_key", sa.String, nullable=False),
-    sa.Column("recorded_at", sa.DateTime, nullable=False),
-    sa.Column(
-        "status",
-        sa.String,
-        nullable=False,
-        server_default="pending",
-    ),
-    sa.Column("meeting_id", sa.String),
-    sa.Index("idx_recording_meeting_id", "meeting_id"),
-)
-
 
 class Recording(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
     id: str = Field(default_factory=generate_uuid4)
-    bucket_name: str
+    meeting_id: str
+    url: str
     object_key: str
-    recorded_at: datetime
-    status: Literal["pending", "processing", "completed", "failed"] = "pending"
-    meeting_id: str | None = None
+    duration: float | None = None
+    created_at: datetime
 
 
 class RecordingController:
-    async def create(self, recording: Recording):
-        query = recordings.insert().values(**recording.model_dump())
-        await database.execute(query)
+    async def create(
+        self,
+        session: AsyncSession,
+        meeting_id: str,
+        url: str,
+        object_key: str,
+        duration: float | None = None,
+        created_at: datetime | None = None,
+    ):
+        if created_at is None:
+            created_at = datetime.now(timezone.utc)
+
+        recording = Recording(
+            meeting_id=meeting_id,
+            url=url,
+            object_key=object_key,
+            duration=duration,
+            created_at=created_at,
+        )
+        new_recording = RecordingModel(**recording.model_dump())
+        session.add(new_recording)
+        await session.commit()
         return recording
 
-    async def get_by_id(self, id: str) -> Recording:
-        query = recordings.select().where(recordings.c.id == id)
-        result = await database.fetch_one(query)
-        return Recording(**result) if result else None
+    async def get_by_id(
+        self, session: AsyncSession, recording_id: str
+    ) -> Recording | None:
+        """
+        Get a recording by id
+        """
+        query = select(RecordingModel).where(RecordingModel.id == recording_id)
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
+            return None
+        return Recording.model_validate(row)
 
-    async def get_by_object_key(self, bucket_name: str, object_key: str) -> Recording:
-        query = recordings.select().where(
-            recordings.c.bucket_name == bucket_name,
-            recordings.c.object_key == object_key,
-        )
-        result = await database.fetch_one(query)
-        return Recording(**result) if result else None
+    async def get_by_meeting_id(
+        self, session: AsyncSession, meeting_id: str
+    ) -> list[Recording]:
+        """
+        Get all recordings for a meeting
+        """
+        query = select(RecordingModel).where(RecordingModel.meeting_id == meeting_id)
+        result = await session.execute(query)
+        return [Recording.model_validate(row) for row in result.scalars().all()]
+
+    async def remove_by_id(self, session: AsyncSession, recording_id: str) -> None:
+        """
+        Remove a recording by id
+        """
+        query = delete(RecordingModel).where(RecordingModel.id == recording_id)
+        await session.execute(query)
+        await session.commit()
 
 
 recordings_controller = RecordingController()

server/reflector/db/rooms.py

@@ -1,54 +1,25 @@
-from datetime import datetime
+import secrets
+from datetime import datetime, timezone
 from sqlite3 import IntegrityError
 from typing import Literal
 
-import sqlalchemy
 from fastapi import HTTPException
-from pydantic import BaseModel, Field
-from sqlalchemy.sql import false, or_
+from pydantic import BaseModel, ConfigDict, Field
+from sqlalchemy import delete, select, update
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.sql import or_
 
-from reflector.db import database, metadata
+from reflector.db.base import RoomModel
 from reflector.utils import generate_uuid4
 
-rooms = sqlalchemy.Table(
-    "room",
-    metadata,
-    sqlalchemy.Column("id", sqlalchemy.String, primary_key=True),
-    sqlalchemy.Column("name", sqlalchemy.String, nullable=False, unique=True),
-    sqlalchemy.Column("user_id", sqlalchemy.String, nullable=False),
-    sqlalchemy.Column("created_at", sqlalchemy.DateTime, nullable=False),
-    sqlalchemy.Column(
-        "zulip_auto_post", sqlalchemy.Boolean, nullable=False, server_default=false()
-    ),
-    sqlalchemy.Column("zulip_stream", sqlalchemy.String),
-    sqlalchemy.Column("zulip_topic", sqlalchemy.String),
-    sqlalchemy.Column(
-        "is_locked", sqlalchemy.Boolean, nullable=False, server_default=false()
-    ),
-    sqlalchemy.Column(
-        "room_mode", sqlalchemy.String, nullable=False, server_default="normal"
-    ),
-    sqlalchemy.Column(
-        "recording_type", sqlalchemy.String, nullable=False, server_default="cloud"
-    ),
-    sqlalchemy.Column(
-        "recording_trigger",
-        sqlalchemy.String,
-        nullable=False,
-        server_default="automatic-2nd-participant",
-    ),
-    sqlalchemy.Column(
-        "is_shared", sqlalchemy.Boolean, nullable=False, server_default=false()
-    ),
-    sqlalchemy.Index("idx_room_is_shared", "is_shared"),
-)
-
 
 class Room(BaseModel):
+    model_config = ConfigDict(from_attributes=True)
+
     id: str = Field(default_factory=generate_uuid4)
     name: str
     user_id: str
-    created_at: datetime = Field(default_factory=datetime.utcnow)
+    created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     zulip_auto_post: bool = False
     zulip_stream: str = ""
     zulip_topic: str = ""
@@ -59,11 +30,19 @@ class Room(BaseModel):
         "none", "prompt", "automatic", "automatic-2nd-participant"
     ] = "automatic-2nd-participant"
     is_shared: bool = False
+    webhook_url: str | None = None
+    webhook_secret: str | None = None
+    ics_url: str | None = None
+    ics_fetch_interval: int = 300
+    ics_enabled: bool = False
+    ics_last_sync: datetime | None = None
+    ics_last_etag: str | None = None
 
 
 class RoomController:
     async def get_all(
         self,
+        session: AsyncSession,
         user_id: str | None = None,
         order_by: str | None = None,
         return_query: bool = False,
@@ -77,14 +56,14 @@ class RoomController:
         Parameters:
         - `order_by`: field to order by, e.g. "-created_at"
         """
-        query = rooms.select()
+        query = select(RoomModel)
         if user_id is not None:
-            query = query.where(or_(rooms.c.user_id == user_id, rooms.c.is_shared))
+            query = query.where(or_(RoomModel.user_id == user_id, RoomModel.is_shared))
         else:
-            query = query.where(rooms.c.is_shared)
+            query = query.where(RoomModel.is_shared)
 
         if order_by is not None:
-            field = getattr(rooms.c, order_by[1:])
+            field = getattr(RoomModel, order_by[1:])
             if order_by.startswith("-"):
                 field = field.desc()
             query = query.order_by(field)
@@ -92,11 +71,12 @@ class RoomController:
         if return_query:
             return query
 
-        results = await database.fetch_all(query)
-        return results
+        result = await session.execute(query)
+        return [Room.model_validate(row) for row in result.scalars().all()]
 
     async def add(
         self,
+        session: AsyncSession,
         name: str,
         user_id: str,
         zulip_auto_post: bool,
@@ -107,10 +87,18 @@ class RoomController:
         recording_type: str,
         recording_trigger: str,
         is_shared: bool,
+        webhook_url: str = "",
+        webhook_secret: str = "",
+        ics_url: str | None = None,
+        ics_fetch_interval: int = 300,
+        ics_enabled: bool = False,
     ):
         """
         Add a new room
         """
+        if webhook_url and not webhook_secret:
+            webhook_secret = secrets.token_urlsafe(32)
+
         room = Room(
             name=name,
             user_id=user_id,
@@ -122,21 +110,33 @@ class RoomController:
             recording_type=recording_type,
             recording_trigger=recording_trigger,
             is_shared=is_shared,
+            webhook_url=webhook_url,
+            webhook_secret=webhook_secret,
+            ics_url=ics_url,
+            ics_fetch_interval=ics_fetch_interval,
+            ics_enabled=ics_enabled,
         )
-        query = rooms.insert().values(**room.model_dump())
+        new_room = RoomModel(**room.model_dump())
+        session.add(new_room)
         try:
-            await database.execute(query)
+            await session.flush()
         except IntegrityError:
             raise HTTPException(status_code=400, detail="Room name is not unique")
         return room
 
-    async def update(self, room: Room, values: dict, mutate=True):
+    async def update(
+        self, session: AsyncSession, room: Room, values: dict, mutate=True
+    ):
         """
         Update a room fields with key/values in values
         """
-        query = rooms.update().where(rooms.c.id == room.id).values(**values)
+        if values.get("webhook_url") and not values.get("webhook_secret"):
+            values["webhook_secret"] = secrets.token_urlsafe(32)
+
+        query = update(RoomModel).where(RoomModel.id == room.id).values(**values)
         try:
-            await database.execute(query)
+            await session.execute(query)
+            await session.flush()
         except IntegrityError:
             raise HTTPException(status_code=400, detail="Room name is not unique")
 
@@ -144,60 +144,79 @@ class RoomController:
             for key, value in values.items():
                 setattr(room, key, value)
 
-    async def get_by_id(self, room_id: str, **kwargs) -> Room | None:
+    async def get_by_id(
+        self, session: AsyncSession, room_id: str, **kwargs
+    ) -> Room | None:
         """
         Get a room by id
         """
-        query = rooms.select().where(rooms.c.id == room_id)
+        query = select(RoomModel).where(RoomModel.id == room_id)
         if "user_id" in kwargs:
-            query = query.where(rooms.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
-        if not result:
+            query = query.where(RoomModel.user_id == kwargs["user_id"])
+        result = await session.execute(query)
+        row = result.scalars().first()
+        if not row:
             return None
-        return Room(**result)
+        return Room.model_validate(row)
 
-    async def get_by_name(self, room_name: str, **kwargs) -> Room | None:
+    async def get_by_name(
+        self, session: AsyncSession, room_name: str, **kwargs
+    ) -> Room | None:
         """
         Get a room by name
         """
-        query = rooms.select().where(rooms.c.name == room_name)
+        query = select(RoomModel).where(RoomModel.name == room_name)
         if "user_id" in kwargs:
-            query = query.where(rooms.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
-        if not result:
+            query = query.where(RoomModel.user_id == kwargs["user_id"])
+        result = await session.execute(query)
+        row = result.scalars().first()
+        if not row:
             return None
-        return Room(**result)
+        return Room.model_validate(row)
 
-    async def get_by_id_for_http(self, meeting_id: str, user_id: str | None) -> Room:
+    async def get_by_id_for_http(
+        self, session: AsyncSession, meeting_id: str, user_id: str | None
+    ) -> Room:
         """
         Get a room by ID for HTTP request.
 
         If not found, it will raise a 404 error.
         """
-        query = rooms.select().where(rooms.c.id == meeting_id)
-        result = await database.fetch_one(query)
-        if not result:
+        query = select(RoomModel).where(RoomModel.id == meeting_id)
+        result = await session.execute(query)
+        row = result.scalars().first()
+        if not row:
             raise HTTPException(status_code=404, detail="Room not found")
 
-        room = Room(**result)
+        room = Room.model_validate(row)
 
         return room
 
+    async def get_ics_enabled(self, session: AsyncSession) -> list[Room]:
+        query = select(RoomModel).where(
+            RoomModel.ics_enabled == True, RoomModel.ics_url != None
+        )
+        result = await session.execute(query)
+        results = result.scalars().all()
+        return [Room(**row.__dict__) for row in results]
+
     async def remove_by_id(
         self,
+        session: AsyncSession,
         room_id: str,
         user_id: str | None = None,
     ) -> None:
         """
         Remove a room by id
         """
-        room = await self.get_by_id(room_id, user_id=user_id)
+        room = await self.get_by_id(session, room_id, user_id=user_id)
         if not room:
             return
         if user_id is not None and room.user_id != user_id:
             return
-        query = rooms.delete().where(rooms.c.id == room_id)
-        await database.execute(query)
+        query = delete(RoomModel).where(RoomModel.id == room_id)
+        await session.execute(query)
+        await session.flush()
 
 
 rooms_controller = RoomController()

server/reflector/db/search.py

@@ -0,0 +1,464 @@
+"""Search functionality for transcripts and other entities."""
+
+import itertools
+from dataclasses import dataclass
+from datetime import datetime
+from io import StringIO
+from typing import Annotated, Any, Dict, Iterator
+
+import sqlalchemy
+import webvtt
+from fastapi import HTTPException
+from pydantic import (
+    BaseModel,
+    Field,
+    NonNegativeFloat,
+    NonNegativeInt,
+    TypeAdapter,
+    ValidationError,
+    constr,
+    field_serializer,
+)
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from reflector.db.base import RoomModel, TranscriptModel
+from reflector.db.transcripts import SourceKind, TranscriptStatus
+from reflector.logger import logger
+from reflector.utils.string import NonEmptyString, try_parse_non_empty_string
+
+DEFAULT_SEARCH_LIMIT = 20
+SNIPPET_CONTEXT_LENGTH = 50  # Characters before/after match to include
+DEFAULT_SNIPPET_MAX_LENGTH = NonNegativeInt(150)
+DEFAULT_MAX_SNIPPETS = NonNegativeInt(3)
+LONG_SUMMARY_MAX_SNIPPETS = 2
+
+SearchQueryBase = constr(min_length=1, strip_whitespace=True)
+SearchLimitBase = Annotated[int, Field(ge=1, le=100)]
+SearchOffsetBase = Annotated[int, Field(ge=0)]
+SearchTotalBase = Annotated[int, Field(ge=0)]
+
+SearchQuery = Annotated[SearchQueryBase, Field(description="Search query text")]
+search_query_adapter = TypeAdapter(SearchQuery)
+SearchLimit = Annotated[SearchLimitBase, Field(description="Results per page")]
+SearchOffset = Annotated[
+    SearchOffsetBase, Field(description="Number of results to skip")
+]
+SearchTotal = Annotated[
+    SearchTotalBase, Field(description="Total number of search results")
+]
+
+WEBVTT_SPEC_HEADER = "WEBVTT"
+
+WebVTTContent = Annotated[
+    str,
+    Field(min_length=len(WEBVTT_SPEC_HEADER), description="WebVTT content"),
+]
+
+
+class WebVTTProcessor:
+    """Stateless processor for WebVTT content operations."""
+
+    @staticmethod
+    def parse(raw_content: str) -> WebVTTContent:
+        """Parse WebVTT content and return it as a string."""
+        if not raw_content.startswith(WEBVTT_SPEC_HEADER):
+            raise ValueError(f"Invalid WebVTT content, no header {WEBVTT_SPEC_HEADER}")
+        return raw_content
+
+    @staticmethod
+    def extract_text(webvtt_content: WebVTTContent) -> str:
+        """Extract plain text from WebVTT content using webvtt library."""
+        try:
+            buffer = StringIO(webvtt_content)
+            vtt = webvtt.read_buffer(buffer)
+            return " ".join(caption.text for caption in vtt if caption.text)
+        except webvtt.errors.MalformedFileError as e:
+            logger.warning(f"Malformed WebVTT content: {e}")
+            return ""
+        except (UnicodeDecodeError, ValueError) as e:
+            logger.warning(f"Failed to decode WebVTT content: {e}")
+            return ""
+        except AttributeError as e:
+            logger.error(
+                f"WebVTT parsing error - unexpected format: {e}", exc_info=True
+            )
+            return ""
+        except Exception as e:
+            logger.error(f"Unexpected error parsing WebVTT: {e}", exc_info=True)
+            return ""
+
+    @staticmethod
+    def generate_snippets(
+        webvtt_content: WebVTTContent,
+        query: SearchQuery,
+        max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
+    ) -> list[str]:
+        """Generate snippets from WebVTT content."""
+        return SnippetGenerator.generate(
+            WebVTTProcessor.extract_text(webvtt_content),
+            query,
+            max_snippets=max_snippets,
+        )
+
+
+@dataclass(frozen=True)
+class SnippetCandidate:
+    """Represents a candidate snippet with its position."""
+
+    _text: str
+    start: NonNegativeInt
+    _original_text_length: int
+
+    @property
+    def end(self) -> NonNegativeInt:
+        """Calculate end position from start and raw text length."""
+        return self.start + len(self._text)
+
+    def text(self) -> str:
+        """Get display text with ellipses added if needed."""
+        result = self._text.strip()
+        if self.start > 0:
+            result = "..." + result
+        if self.end < self._original_text_length:
+            result = result + "..."
+        return result
+
+
+class SearchParameters(BaseModel):
+    """Validated search parameters for full-text search."""
+
+    query_text: SearchQuery | None = None
+    limit: SearchLimit = DEFAULT_SEARCH_LIMIT
+    offset: SearchOffset = 0
+    user_id: str | None = None
+    room_id: str | None = None
+    source_kind: SourceKind | None = None
+
+
+class SearchResultDB(BaseModel):
+    """Intermediate model for validating raw database results."""
+
+    id: str = Field(..., min_length=1)
+    created_at: datetime
+    status: str = Field(..., min_length=1)
+    duration: float | None = Field(None, ge=0)
+    user_id: str | None = None
+    title: str | None = None
+    source_kind: SourceKind
+    room_id: str | None = None
+    rank: float = Field(..., ge=0, le=1)
+
+
+class SearchResult(BaseModel):
+    """Public search result model with computed fields."""
+
+    id: str = Field(..., min_length=1)
+    title: str | None = None
+    user_id: str | None = None
+    room_id: str | None = None
+    room_name: str | None = None
+    source_kind: SourceKind
+    created_at: datetime
+    status: TranscriptStatus = Field(..., min_length=1)
+    rank: float = Field(..., ge=0, le=1)
+    duration: NonNegativeFloat | None = Field(..., description="Duration in seconds")
+    search_snippets: list[str] = Field(
+        description="Text snippets around search matches"
+    )
+    total_match_count: NonNegativeInt = Field(
+        default=0, description="Total number of matches found in the transcript"
+    )
+
+    @field_serializer("created_at", when_used="json")
+    def serialize_datetime(self, dt: datetime) -> str:
+        if dt.tzinfo is None:
+            return dt.isoformat() + "Z"
+        return dt.isoformat()
+
+
+class SnippetGenerator:
+    """Stateless generator for text snippets and match operations."""
+
+    @staticmethod
+    def find_all_matches(text: str, query: str) -> Iterator[int]:
+        """Generate all match positions for a query in text."""
+        if not text:
+            logger.warning("Empty text for search query in find_all_matches")
+            return
+        if not query:
+            logger.warning("Empty query for search text in find_all_matches")
+            return
+
+        text_lower = text.lower()
+        query_lower = query.lower()
+        start = 0
+        prev_start = start
+        while (pos := text_lower.find(query_lower, start)) != -1:
+            yield pos
+            start = pos + len(query_lower)
+            if start <= prev_start:
+                raise ValueError("panic! find_all_matches is not incremental")
+            prev_start = start
+
+    @staticmethod
+    def count_matches(text: str, query: SearchQuery) -> NonNegativeInt:
+        """Count total number of matches for a query in text."""
+        ZERO = NonNegativeInt(0)
+        if not text:
+            logger.warning("Empty text for search query in count_matches")
+            return ZERO
+        assert query is not None
+        return NonNegativeInt(
+            sum(1 for _ in SnippetGenerator.find_all_matches(text, query))
+        )
+
+    @staticmethod
+    def create_snippet(
+        text: str, match_pos: int, max_length: int = DEFAULT_SNIPPET_MAX_LENGTH
+    ) -> SnippetCandidate:
+        """Create a snippet from a match position."""
+        snippet_start = NonNegativeInt(max(0, match_pos - SNIPPET_CONTEXT_LENGTH))
+        snippet_end = min(len(text), match_pos + max_length - SNIPPET_CONTEXT_LENGTH)
+
+        snippet_text = text[snippet_start:snippet_end]
+
+        return SnippetCandidate(
+            _text=snippet_text, start=snippet_start, _original_text_length=len(text)
+        )
+
+    @staticmethod
+    def filter_non_overlapping(
+        candidates: Iterator[SnippetCandidate],
+    ) -> Iterator[str]:
+        """Filter out overlapping snippets and return only display text."""
+        last_end = 0
+        for candidate in candidates:
+            display_text = candidate.text()
+            # it means that next overlapping snippets simply don't get included
+            # it's fine as simplistic logic and users probably won't care much because they already have their search results just fin
+            if candidate.start >= last_end and display_text:
+                yield display_text
+                last_end = candidate.end
+
+    @staticmethod
+    def generate(
+        text: str,
+        query: SearchQuery,
+        max_length: NonNegativeInt = DEFAULT_SNIPPET_MAX_LENGTH,
+        max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
+    ) -> list[str]:
+        """Generate snippets from text."""
+        assert query is not None
+        if not text:
+            logger.warning("Empty text for generate_snippets")
+            return []
+
+        candidates = (
+            SnippetGenerator.create_snippet(text, pos, max_length)
+            for pos in SnippetGenerator.find_all_matches(text, query)
+        )
+        filtered = SnippetGenerator.filter_non_overlapping(candidates)
+        snippets = list(itertools.islice(filtered, max_snippets))
+
+        # Fallback to first word search if no full matches
+        # it's another assumption: proper snippet logic generation is quite complicated and tied to db logic, so simplification is used here
+        if not snippets and " " in query:
+            first_word = query.split()[0]
+            return SnippetGenerator.generate(text, first_word, max_length, max_snippets)
+
+        return snippets
+
+    @staticmethod
+    def from_summary(
+        summary: str,
+        query: SearchQuery,
+        max_snippets: NonNegativeInt = LONG_SUMMARY_MAX_SNIPPETS,
+    ) -> list[str]:
+        """Generate snippets from summary text."""
+        return SnippetGenerator.generate(summary, query, max_snippets=max_snippets)
+
+    @staticmethod
+    def combine_sources(
+        summary: NonEmptyString | None,
+        webvtt: WebVTTContent | None,
+        query: SearchQuery,
+        max_total: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
+    ) -> tuple[list[str], NonNegativeInt]:
+        """Combine snippets from multiple sources and return total match count.
+
+        Returns (snippets, total_match_count) tuple.
+
+        snippets can be empty for real in case of e.g. title match
+        """
+
+        assert (
+            summary is not None or webvtt is not None
+        ), "At least one source must be present"
+
+        webvtt_matches = 0
+        summary_matches = 0
+
+        if webvtt:
+            webvtt_text = WebVTTProcessor.extract_text(webvtt)
+            webvtt_matches = SnippetGenerator.count_matches(webvtt_text, query)
+
+        if summary:
+            summary_matches = SnippetGenerator.count_matches(summary, query)
+
+        total_matches = NonNegativeInt(webvtt_matches + summary_matches)
+
+        summary_snippets = (
+            SnippetGenerator.from_summary(summary, query) if summary else []
+        )
+
+        if len(summary_snippets) >= max_total:
+            return summary_snippets[:max_total], total_matches
+
+        remaining = max_total - len(summary_snippets)
+        webvtt_snippets = (
+            WebVTTProcessor.generate_snippets(webvtt, query, remaining)
+            if webvtt
+            else []
+        )
+
+        return summary_snippets + webvtt_snippets, total_matches
+
+
+class SearchController:
+    """Controller for search operations across different entities."""
+
+    @classmethod
+    async def search_transcripts(
+        cls, session: AsyncSession, params: SearchParameters
+    ) -> tuple[list[SearchResult], int]:
+        """
+        Full-text search for transcripts using PostgreSQL tsvector.
+        Returns (results, total_count).
+        """
+
+        base_columns = [
+            TranscriptModel.id,
+            TranscriptModel.title,
+            TranscriptModel.created_at,
+            TranscriptModel.duration,
+            TranscriptModel.status,
+            TranscriptModel.user_id,
+            TranscriptModel.room_id,
+            TranscriptModel.source_kind,
+            TranscriptModel.webvtt,
+            TranscriptModel.long_summary,
+            sqlalchemy.case(
+                (
+                    TranscriptModel.room_id.isnot(None) & RoomModel.id.is_(None),
+                    "Deleted Room",
+                ),
+                else_=RoomModel.name,
+            ).label("room_name"),
+        ]
+        search_query = None
+        if params.query_text is not None:
+            search_query = sqlalchemy.func.websearch_to_tsquery(
+                "english", params.query_text
+            )
+            rank_column = sqlalchemy.func.ts_rank(
+                TranscriptModel.search_vector_en,
+                search_query,
+                32,  # normalization flag: rank/(rank+1) for 0-1 range
+            ).label("rank")
+        else:
+            rank_column = sqlalchemy.cast(1.0, sqlalchemy.Float).label("rank")
+
+        columns = base_columns + [rank_column]
+        base_query = (
+            sqlalchemy.select(*columns)
+            .select_from(TranscriptModel)
+            .outerjoin(RoomModel, TranscriptModel.room_id == RoomModel.id)
+        )
+
+        if params.query_text is not None:
+            # because already initialized based on params.query_text presence above
+            assert search_query is not None
+            base_query = base_query.where(
+                TranscriptModel.search_vector_en.op("@@")(search_query)
+            )
+
+        if params.user_id:
+            base_query = base_query.where(
+                sqlalchemy.or_(
+                    TranscriptModel.user_id == params.user_id, RoomModel.is_shared
+                )
+            )
+        else:
+            base_query = base_query.where(RoomModel.is_shared)
+        if params.room_id:
+            base_query = base_query.where(TranscriptModel.room_id == params.room_id)
+        if params.source_kind:
+            base_query = base_query.where(
+                TranscriptModel.source_kind == params.source_kind
+            )
+
+        if params.query_text is not None:
+            order_by = sqlalchemy.desc(sqlalchemy.text("rank"))
+        else:
+            order_by = sqlalchemy.desc(TranscriptModel.created_at)
+
+        query = base_query.order_by(order_by).limit(params.limit).offset(params.offset)
+
+        result = await session.execute(query)
+        rs = result.mappings().all()
+
+        count_query = sqlalchemy.select(sqlalchemy.func.count()).select_from(
+            base_query.alias("search_results")
+        )
+        count_result = await session.execute(count_query)
+        total = count_result.scalar()
+
+        def _process_result(r: dict) -> SearchResult:
+            r_dict: Dict[str, Any] = dict(r)
+
+            webvtt_raw: str | None = r_dict.pop("webvtt", None)
+            webvtt: WebVTTContent | None
+            if webvtt_raw:
+                webvtt = WebVTTProcessor.parse(webvtt_raw)
+            else:
+                webvtt = None
+
+            long_summary_r: str | None = r_dict.pop("long_summary", None)
+            long_summary: NonEmptyString = try_parse_non_empty_string(long_summary_r)
+            room_name: str | None = r_dict.pop("room_name", None)
+            db_result = SearchResultDB.model_validate(r_dict)
+
+            at_least_one_source = webvtt is not None or long_summary is not None
+            has_query = params.query_text is not None
+            snippets, total_match_count = (
+                SnippetGenerator.combine_sources(
+                    long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+                )
+                if has_query and at_least_one_source
+                else ([], 0)
+            )
+
+            return SearchResult(
+                **db_result.model_dump(),
+                room_name=room_name,
+                search_snippets=snippets,
+                total_match_count=total_match_count,
+            )
+
+        try:
+            results = [_process_result(r) for r in rs]
+        except ValidationError as e:
+            logger.error(f"Invalid search result data: {e}", exc_info=True)
+            raise HTTPException(
+                status_code=500, detail="Internal search result data consistency error"
+            )
+        except Exception as e:
+            logger.error(f"Error processing search results: {e}", exc_info=True)
+            raise
+
+        return results, total
+
+
+search_controller = SearchController()
+webvtt_processor = WebVTTProcessor()
+snippet_generator = SnippetGenerator()

server/reflector/db/transcripts.py

@@ -3,21 +3,24 @@ import json
 import os
 import shutil
 from contextlib import asynccontextmanager
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 from pathlib import Path
 from typing import Any, Literal
 
-import sqlalchemy
 from fastapi import HTTPException
 from pydantic import BaseModel, ConfigDict, Field, field_serializer
-from sqlalchemy import Enum
-from sqlalchemy.sql import false, or_
+from sqlalchemy import delete, insert, select, update
+from sqlalchemy.ext.asyncio import AsyncSession
+from sqlalchemy.sql import or_
 
-from reflector.db import database, metadata
+from reflector.db.base import RoomModel, TranscriptModel
+from reflector.db.recordings import recordings_controller
+from reflector.logger import logger
 from reflector.processors.types import Word as ProcessorWord
 from reflector.settings import settings
-from reflector.storage import get_transcripts_storage
+from reflector.storage import get_recordings_storage, get_transcripts_storage
 from reflector.utils import generate_uuid4
+from reflector.utils.webvtt import topics_to_webvtt
 
 
 class SourceKind(enum.StrEnum):
@@ -26,69 +29,20 @@ class SourceKind(enum.StrEnum):
     FILE = enum.auto()
 
 
-transcripts = sqlalchemy.Table(
-    "transcript",
-    metadata,
-    sqlalchemy.Column("id", sqlalchemy.String, primary_key=True),
-    sqlalchemy.Column("name", sqlalchemy.String),
-    sqlalchemy.Column("status", sqlalchemy.String),
-    sqlalchemy.Column("locked", sqlalchemy.Boolean),
-    sqlalchemy.Column("duration", sqlalchemy.Float),
-    sqlalchemy.Column("created_at", sqlalchemy.DateTime),
-    sqlalchemy.Column("title", sqlalchemy.String),
-    sqlalchemy.Column("short_summary", sqlalchemy.String),
-    sqlalchemy.Column("long_summary", sqlalchemy.String),
-    sqlalchemy.Column("topics", sqlalchemy.JSON),
-    sqlalchemy.Column("events", sqlalchemy.JSON),
-    sqlalchemy.Column("participants", sqlalchemy.JSON),
-    sqlalchemy.Column("source_language", sqlalchemy.String),
-    sqlalchemy.Column("target_language", sqlalchemy.String),
-    sqlalchemy.Column(
-        "reviewed", sqlalchemy.Boolean, nullable=False, server_default=false()
-    ),
-    sqlalchemy.Column(
-        "audio_location",
-        sqlalchemy.String,
-        nullable=False,
-        server_default="local",
-    ),
-    # with user attached, optional
-    sqlalchemy.Column("user_id", sqlalchemy.String),
-    sqlalchemy.Column(
-        "share_mode",
-        sqlalchemy.String,
-        nullable=False,
-        server_default="private",
-    ),
-    sqlalchemy.Column(
-        "meeting_id",
-        sqlalchemy.String,
-    ),
-    sqlalchemy.Column("recording_id", sqlalchemy.String),
-    sqlalchemy.Column("zulip_message_id", sqlalchemy.Integer),
-    sqlalchemy.Column(
-        "source_kind",
-        Enum(SourceKind, values_callable=lambda obj: [e.value for e in obj]),
-        nullable=False,
-    ),
-    # indicative field: whether associated audio is deleted
-    # the main "audio deleted" is the presence of the audio itself / consents not-given
-    # same field could've been in recording/meeting, and it's maybe even ok to dupe it at need
-    sqlalchemy.Column("audio_deleted", sqlalchemy.Boolean),
-    sqlalchemy.Column("room_id", sqlalchemy.String),
-    sqlalchemy.Index("idx_transcript_recording_id", "recording_id"),
-    sqlalchemy.Index("idx_transcript_user_id", "user_id"),
-    sqlalchemy.Index("idx_transcript_created_at", "created_at"),
-    sqlalchemy.Index("idx_transcript_user_id_recording_id", "user_id", "recording_id"),
-    sqlalchemy.Index("idx_transcript_room_id", "room_id"),
-)
-
-
 def generate_transcript_name() -> str:
     now = datetime.now(timezone.utc)
     return f"Transcript {now.strftime('%Y-%m-%d %H:%M:%S')}"
 
 
+TranscriptStatus = Literal[
+    "idle", "uploaded", "recording", "processing", "error", "ended"
+]
+
+
+class StrValue(BaseModel):
+    value: str
+
+
 class AudioWaveform(BaseModel):
     data: list[float]
 
@@ -147,14 +101,20 @@ class TranscriptParticipant(BaseModel):
 
 
 class Transcript(BaseModel):
+    """Full transcript model with all fields."""
+
+    model_config = ConfigDict(from_attributes=True)
+
     id: str = Field(default_factory=generate_uuid4)
     user_id: str | None = None
     name: str = Field(default_factory=generate_transcript_name)
-    status: str = "idle"
-    locked: bool = False
+    status: TranscriptStatus = "idle"
     duration: float = 0
     created_at: datetime = Field(default_factory=lambda: datetime.now(timezone.utc))
     title: str | None = None
+    source_kind: SourceKind
+    room_id: str | None = None
+    locked: bool = False
     short_summary: str | None = None
     long_summary: str | None = None
     topics: list[TranscriptTopic] = []
@@ -168,9 +128,8 @@ class Transcript(BaseModel):
     meeting_id: str | None = None
     recording_id: str | None = None
     zulip_message_id: int | None = None
-    source_kind: SourceKind
     audio_deleted: bool | None = None
-    room_id: str | None = None
+    webvtt: str | None = None
 
     @field_serializer("created_at", when_used="json")
     def serialize_datetime(self, dt: datetime) -> str:
@@ -271,10 +230,12 @@ class Transcript(BaseModel):
         # we need to create an url to be used for diarization
         # we can't use the audio_mp3_filename because it's not accessible
         # from the diarization processor
-        from datetime import timedelta
 
-        from reflector.app import app
-        from reflector.views.transcripts import create_access_token
+        # TODO don't import app in db
+        from reflector.app import app  # noqa: PLC0415
+
+        # TODO a util + don''t import views in db
+        from reflector.views.transcripts import create_access_token  # noqa: PLC0415
 
         path = app.url_path_for(
             "transcript_get_audio_mp3",
@@ -312,6 +273,7 @@ class Transcript(BaseModel):
 class TranscriptController:
     async def get_all(
         self,
+        session: AsyncSession,
         user_id: str | None = None,
         order_by: str | None = None,
         filter_empty: bool | None = False,
@@ -335,104 +297,115 @@ class TranscriptController:
         - `room_id`: filter transcripts by room ID
         - `search_term`: filter transcripts by search term
         """
-        from reflector.db.rooms import rooms
 
-        query = transcripts.select().join(
-            rooms, transcripts.c.room_id == rooms.c.id, isouter=True
+        query = select(TranscriptModel).join(
+            RoomModel, TranscriptModel.room_id == RoomModel.id, isouter=True
         )
 
         if user_id:
             query = query.where(
-                or_(transcripts.c.user_id == user_id, rooms.c.is_shared)
+                or_(TranscriptModel.user_id == user_id, RoomModel.is_shared)
             )
         else:
-            query = query.where(rooms.c.is_shared)
+            query = query.where(RoomModel.is_shared)
 
         if source_kind:
-            query = query.where(transcripts.c.source_kind == source_kind)
+            query = query.where(TranscriptModel.source_kind == source_kind)
 
         if room_id:
-            query = query.where(transcripts.c.room_id == room_id)
+            query = query.where(TranscriptModel.room_id == room_id)
 
         if search_term:
-            query = query.where(transcripts.c.title.ilike(f"%{search_term}%"))
+            query = query.where(TranscriptModel.title.ilike(f"%{search_term}%"))
 
         # Exclude heavy JSON columns from list queries
+        # Get all ORM column attributes except excluded ones
         transcript_columns = [
-            col for col in transcripts.c if col.name not in exclude_columns
+            getattr(TranscriptModel, col.name)
+            for col in TranscriptModel.__table__.c
+            if col.name not in exclude_columns
         ]
 
         query = query.with_only_columns(
-            transcript_columns
-            + [
-                rooms.c.name.label("room_name"),
-            ]
+            *transcript_columns,
+            RoomModel.name.label("room_name"),
         )
 
         if order_by is not None:
-            field = getattr(transcripts.c, order_by[1:])
+            field = getattr(TranscriptModel, order_by[1:])
             if order_by.startswith("-"):
                 field = field.desc()
             query = query.order_by(field)
 
         if filter_empty:
-            query = query.filter(transcripts.c.status != "idle")
+            query = query.filter(TranscriptModel.status != "idle")
 
         if filter_recording:
-            query = query.filter(transcripts.c.status != "recording")
+            query = query.filter(TranscriptModel.status != "recording")
 
         # print(query.compile(compile_kwargs={"literal_binds": True}))
 
         if return_query:
             return query
 
-        results = await database.fetch_all(query)
-        return results
+        result = await session.execute(query)
+        return [dict(row) for row in result.mappings().all()]
 
-    async def get_by_id(self, transcript_id: str, **kwargs) -> Transcript | None:
+    async def get_by_id(
+        self, session: AsyncSession, transcript_id: str, **kwargs
+    ) -> Transcript | None:
         """
         Get a transcript by id
         """
-        query = transcripts.select().where(transcripts.c.id == transcript_id)
+        query = select(TranscriptModel).where(TranscriptModel.id == transcript_id)
         if "user_id" in kwargs:
-            query = query.where(transcripts.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
-        if not result:
+            query = query.where(TranscriptModel.user_id == kwargs["user_id"])
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             return None
-        return Transcript(**result)
+        return Transcript.model_validate(row)
 
     async def get_by_recording_id(
-        self, recording_id: str, **kwargs
+        self, session: AsyncSession, recording_id: str, **kwargs
     ) -> Transcript | None:
         """
         Get a transcript by recording_id
         """
-        query = transcripts.select().where(transcripts.c.recording_id == recording_id)
+        query = select(TranscriptModel).where(
+            TranscriptModel.recording_id == recording_id
+        )
         if "user_id" in kwargs:
-            query = query.where(transcripts.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
-        if not result:
+            query = query.where(TranscriptModel.user_id == kwargs["user_id"])
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             return None
-        return Transcript(**result)
+        return Transcript.model_validate(row)
 
-    async def get_by_room_id(self, room_id: str, **kwargs) -> list[Transcript]:
+    async def get_by_room_id(
+        self, session: AsyncSession, room_id: str, **kwargs
+    ) -> list[Transcript]:
         """
         Get transcripts by room_id (direct access without joins)
         """
-        query = transcripts.select().where(transcripts.c.room_id == room_id)
+        query = select(TranscriptModel).where(TranscriptModel.room_id == room_id)
         if "user_id" in kwargs:
-            query = query.where(transcripts.c.user_id == kwargs["user_id"])
+            query = query.where(TranscriptModel.user_id == kwargs["user_id"])
         if "order_by" in kwargs:
             order_by = kwargs["order_by"]
-            field = getattr(transcripts.c, order_by[1:])
+            field = getattr(TranscriptModel, order_by[1:])
             if order_by.startswith("-"):
                 field = field.desc()
             query = query.order_by(field)
-        results = await database.fetch_all(query)
-        return [Transcript(**result) for result in results]
+        results = await session.execute(query)
+        return [
+            Transcript.model_validate(dict(row)) for row in results.mappings().all()
+        ]
 
     async def get_by_id_for_http(
         self,
+        session: AsyncSession,
         transcript_id: str,
         user_id: str | None,
     ) -> Transcript:
@@ -445,13 +418,14 @@ class TranscriptController:
         This method checks the share mode of the transcript and the user_id
         to determine if the user can access the transcript.
         """
-        query = transcripts.select().where(transcripts.c.id == transcript_id)
-        result = await database.fetch_one(query)
-        if not result:
+        query = select(TranscriptModel).where(TranscriptModel.id == transcript_id)
+        result = await session.execute(query)
+        row = result.scalar_one_or_none()
+        if not row:
             raise HTTPException(status_code=404, detail="Transcript not found")
 
         # if the transcript is anonymous, share mode is not checked
-        transcript = Transcript(**result)
+        transcript = Transcript.model_validate(row)
         if transcript.user_id is None:
             return transcript
 
@@ -474,6 +448,7 @@ class TranscriptController:
 
     async def add(
         self,
+        session: AsyncSession,
         name: str,
         source_kind: SourceKind,
         source_language: str = "en",
@@ -498,58 +473,129 @@ class TranscriptController:
             meeting_id=meeting_id,
             room_id=room_id,
         )
-        query = transcripts.insert().values(**transcript.model_dump())
-        await database.execute(query)
+        query = insert(TranscriptModel).values(**transcript.model_dump())
+        await session.execute(query)
+        await session.commit()
         return transcript
 
-    async def update(self, transcript: Transcript, values: dict, mutate=True):
+    # TODO investigate why mutate= is used. it's used in one place currently, maybe because of ORM field updates.
+    # using mutate=True is discouraged
+    async def update(
+        self, session: AsyncSession, transcript: Transcript, values: dict, mutate=False
+    ) -> Transcript:
         """
-        Update a transcript fields with key/values in values
+        Update a transcript fields with key/values in values.
+        Returns a copy of the transcript with updated values.
         """
+        values = TranscriptController._handle_topics_update(values)
+
         query = (
-            transcripts.update()
-            .where(transcripts.c.id == transcript.id)
+            update(TranscriptModel)
+            .where(TranscriptModel.id == transcript.id)
             .values(**values)
         )
-        await database.execute(query)
+        await session.execute(query)
+        await session.commit()
         if mutate:
             for key, value in values.items():
                 setattr(transcript, key, value)
 
+        updated_transcript = transcript.model_copy(update=values)
+        return updated_transcript
+
+    @staticmethod
+    def _handle_topics_update(values: dict) -> dict:
+        """Auto-update WebVTT when topics are updated."""
+
+        if values.get("webvtt") is not None:
+            logger.warn("trying to update read-only webvtt column")
+            pass
+
+        topics_data = values.get("topics")
+        if topics_data is None:
+            return values
+
+        return {
+            **values,
+            "webvtt": topics_to_webvtt(
+                [TranscriptTopic(**topic_dict) for topic_dict in topics_data]
+            ),
+        }
+
     async def remove_by_id(
         self,
+        session: AsyncSession,
         transcript_id: str,
         user_id: str | None = None,
     ) -> None:
         """
         Remove a transcript by id
         """
-        transcript = await self.get_by_id(transcript_id)
+        transcript = await self.get_by_id(session, transcript_id)
         if not transcript:
             return
         if user_id is not None and transcript.user_id != user_id:
             return
+        if transcript.audio_location == "storage" and not transcript.audio_deleted:
+            try:
+                await get_transcripts_storage().delete_file(
+                    transcript.storage_audio_path
+                )
+            except Exception as e:
+                logger.warning(
+                    "Failed to delete transcript audio from storage",
+                    exc_info=e,
+                    transcript_id=transcript.id,
+                )
         transcript.unlink()
-        query = transcripts.delete().where(transcripts.c.id == transcript_id)
-        await database.execute(query)
+        if transcript.recording_id:
+            try:
+                recording = await recordings_controller.get_by_id(
+                    session, transcript.recording_id
+                )
+                if recording:
+                    try:
+                        await get_recordings_storage().delete_file(recording.object_key)
+                    except Exception as e:
+                        logger.warning(
+                            "Failed to delete recording object from S3",
+                            exc_info=e,
+                            recording_id=transcript.recording_id,
+                        )
+                    await recordings_controller.remove_by_id(
+                        session, transcript.recording_id
+                    )
+            except Exception as e:
+                logger.warning(
+                    "Failed to delete recording row",
+                    exc_info=e,
+                    recording_id=transcript.recording_id,
+                )
+        query = delete(TranscriptModel).where(TranscriptModel.id == transcript_id)
+        await session.execute(query)
+        await session.commit()
 
-    async def remove_by_recording_id(self, recording_id: str):
+    async def remove_by_recording_id(self, session: AsyncSession, recording_id: str):
         """
         Remove a transcript by recording_id
         """
-        query = transcripts.delete().where(transcripts.c.recording_id == recording_id)
-        await database.execute(query)
+        query = delete(TranscriptModel).where(
+            TranscriptModel.recording_id == recording_id
+        )
+        await session.execute(query)
+        await session.commit()
 
     @asynccontextmanager
-    async def transaction(self):
+    async def transaction(self, session: AsyncSession):
         """
         A context manager for database transaction
         """
-        async with database.transaction(isolation="serializable"):
+        async with session.begin():
             yield
 
     async def append_event(
         self,
+        session: AsyncSession,
         transcript: Transcript,
         event: str,
         data: Any,
@@ -558,15 +604,12 @@ class TranscriptController:
         Append an event to a transcript
         """
         resp = transcript.add_event(event=event, data=data)
-        await self.update(
-            transcript,
-            {"events": transcript.events_dump()},
-            mutate=False,
-        )
+        await self.update(session, transcript, {"events": transcript.events_dump()})
         return resp
 
     async def upsert_topic(
         self,
+        session: AsyncSession,
         transcript: Transcript,
         topic: TranscriptTopic,
     ) -> TranscriptEvent:
@@ -574,13 +617,9 @@ class TranscriptController:
         Upsert topics to a transcript
         """
         transcript.upsert_topic(topic)
-        await self.update(
-            transcript,
-            {"topics": transcript.topics_dump()},
-            mutate=False,
-        )
+        await self.update(session, transcript, {"topics": transcript.topics_dump()})
 
-    async def move_mp3_to_storage(self, transcript: Transcript):
+    async def move_mp3_to_storage(self, session: AsyncSession, transcript: Transcript):
         """
         Move mp3 file to storage
         """
@@ -603,12 +642,17 @@ class TranscriptController:
             )
 
             # indicate on the transcript that the audio is now on storage
-            await self.update(transcript, {"audio_location": "storage"})
+            # mutates transcript argument
+            await self.update(
+                session, transcript, {"audio_location": "storage"}, mutate=True
+            )
 
         # unlink the local file
         transcript.audio_mp3_filename.unlink(missing_ok=True)
 
-    async def download_mp3_from_storage(self, transcript: Transcript):
+    async def download_mp3_from_storage(
+        self, session: AsyncSession, transcript: Transcript
+    ):
         """
         Download audio from storage
         """
@@ -620,6 +664,7 @@ class TranscriptController:
 
     async def upsert_participant(
         self,
+        session: AsyncSession,
         transcript: Transcript,
         participant: TranscriptParticipant,
     ) -> TranscriptParticipant:
@@ -628,14 +673,13 @@ class TranscriptController:
         """
         result = transcript.upsert_participant(participant)
         await self.update(
-            transcript,
-            {"participants": transcript.participants_dump()},
-            mutate=False,
+            session, transcript, {"participants": transcript.participants_dump()}
         )
         return result
 
     async def delete_participant(
         self,
+        session: AsyncSession,
         transcript: Transcript,
         participant_id: str,
     ):
@@ -644,10 +688,31 @@ class TranscriptController:
         """
         transcript.delete_participant(participant_id)
         await self.update(
-            transcript,
-            {"participants": transcript.participants_dump()},
-            mutate=False,
+            session, transcript, {"participants": transcript.participants_dump()}
         )
 
+    async def set_status(
+        self, session: AsyncSession, transcript_id: str, status: TranscriptStatus
+    ) -> TranscriptEvent | None:
+        """
+        Update the status of a transcript
+
+        Will add an event STATUS + update the status field of transcript
+        """
+        async with self.transaction(session):
+            transcript = await self.get_by_id(session, transcript_id)
+            if not transcript:
+                raise Exception(f"Transcript {transcript_id} not found")
+            if transcript.status == status:
+                return
+            resp = await self.append_event(
+                session,
+                transcript=transcript,
+                event="STATUS",
+                data=StrValue(value=status),
+            )
+            await self.update(session, transcript, {"status": status})
+        return resp
+
 
 transcripts_controller = TranscriptController()

server/reflector/llm.py

@@ -0,0 +1,83 @@
+from typing import Type, TypeVar
+
+from llama_index.core import Settings
+from llama_index.core.output_parsers import PydanticOutputParser
+from llama_index.core.program import LLMTextCompletionProgram
+from llama_index.core.response_synthesizers import TreeSummarize
+from llama_index.llms.openai_like import OpenAILike
+from pydantic import BaseModel
+
+T = TypeVar("T", bound=BaseModel)
+
+STRUCTURED_RESPONSE_PROMPT_TEMPLATE = """
+Based on the following analysis, provide the information in the requested JSON format:
+
+Analysis:
+{analysis}
+
+{format_instructions}
+"""
+
+
+class LLM:
+    def __init__(self, settings, temperature: float = 0.4, max_tokens: int = 2048):
+        self.settings_obj = settings
+        self.model_name = settings.LLM_MODEL
+        self.url = settings.LLM_URL
+        self.api_key = settings.LLM_API_KEY
+        self.context_window = settings.LLM_CONTEXT_WINDOW
+        self.temperature = temperature
+        self.max_tokens = max_tokens
+
+        # Configure llamaindex Settings
+        self._configure_llamaindex()
+
+    def _configure_llamaindex(self):
+        """Configure llamaindex Settings with OpenAILike LLM"""
+        Settings.llm = OpenAILike(
+            model=self.model_name,
+            api_base=self.url,
+            api_key=self.api_key,
+            context_window=self.context_window,
+            is_chat_model=True,
+            is_function_calling_model=False,
+            temperature=self.temperature,
+            max_tokens=self.max_tokens,
+        )
+
+    async def get_response(
+        self, prompt: str, texts: list[str], tone_name: str | None = None
+    ) -> str:
+        """Get a text response using TreeSummarize for non-function-calling models"""
+        summarizer = TreeSummarize(verbose=False)
+        response = await summarizer.aget_response(prompt, texts, tone_name=tone_name)
+        return str(response).strip()
+
+    async def get_structured_response(
+        self,
+        prompt: str,
+        texts: list[str],
+        output_cls: Type[T],
+        tone_name: str | None = None,
+    ) -> T:
+        """Get structured output from LLM for non-function-calling models"""
+        summarizer = TreeSummarize(verbose=True)
+        response = await summarizer.aget_response(prompt, texts, tone_name=tone_name)
+
+        output_parser = PydanticOutputParser(output_cls)
+
+        program = LLMTextCompletionProgram.from_defaults(
+            output_parser=output_parser,
+            prompt_template_str=STRUCTURED_RESPONSE_PROMPT_TEMPLATE,
+            verbose=False,
+        )
+
+        format_instructions = output_parser.format(
+            "Please structure the above information in the following JSON format:"
+        )
+
+        output = await program.acall(
+            analysis=str(response), format_instructions=format_instructions
+        )
+
+        return output

server/reflector/llm/__init__.py

@@ -1,2 +0,0 @@
-from .base import LLM  # noqa: F401
-from .llm_params import LLMTaskParams  # noqa: F401

server/reflector/llm/base.py

@@ -1,347 +0,0 @@
-import importlib
-import json
-import re
-from typing import TypeVar
-
-import nltk
-from prometheus_client import Counter, Histogram
-from transformers import GenerationConfig
-
-from reflector.llm.llm_params import TaskParams
-from reflector.logger import logger as reflector_logger
-from reflector.settings import settings
-from reflector.utils.retry import retry
-
-T = TypeVar("T", bound="LLM")
-
-
-class LLM:
-    _nltk_downloaded = False
-    _registry = {}
-    model_name: str
-    m_generate = Histogram(
-        "llm_generate",
-        "Time spent in LLM.generate",
-        ["backend"],
-    )
-    m_generate_call = Counter(
-        "llm_generate_call",
-        "Number of calls to LLM.generate",
-        ["backend"],
-    )
-    m_generate_success = Counter(
-        "llm_generate_success",
-        "Number of successful calls to LLM.generate",
-        ["backend"],
-    )
-    m_generate_failure = Counter(
-        "llm_generate_failure",
-        "Number of failed calls to LLM.generate",
-        ["backend"],
-    )
-
-    @classmethod
-    def ensure_nltk(cls):
-        """
-        Make sure NLTK package is installed. Searches in the cache and
-        downloads only if needed.
-        """
-        if not cls._nltk_downloaded:
-            nltk.download("punkt_tab")
-            # For POS tagging
-            nltk.download("averaged_perceptron_tagger_eng")
-            cls._nltk_downloaded = True
-
-    @classmethod
-    def register(cls, name, klass):
-        cls._registry[name] = klass
-
-    @classmethod
-    def get_instance(cls, model_name: str | None = None, name: str = None) -> T:
-        """
-        Return an instance depending on the settings.
-        Settings used:
-
-        - `LLM_BACKEND`: key of the backend
-        - `LLM_URL`: url of the backend
-        """
-        if name is None:
-            name = settings.LLM_BACKEND
-        if name not in cls._registry:
-            module_name = f"reflector.llm.llm_{name}"
-            importlib.import_module(module_name)
-        cls.ensure_nltk()
-
-        return cls._registry[name](model_name)
-
-    def get_model_name(self) -> str:
-        """
-        Get the currently set model name
-        """
-        return self._get_model_name()
-
-    def _get_model_name(self) -> str:
-        pass
-
-    def set_model_name(self, model_name: str) -> bool:
-        """
-        Update the model name with the provided model name
-        """
-        return self._set_model_name(model_name)
-
-    def _set_model_name(self, model_name: str) -> bool:
-        raise NotImplementedError
-
-    @property
-    def template(self) -> str:
-        """
-        Return the LLM Prompt template
-        """
-        return """
-        ### Human:
-        {instruct}
-
-        {text}
-
-        ### Assistant:
-        """
-
-    def __init__(self):
-        name = self.__class__.__name__
-        self.m_generate = self.m_generate.labels(name)
-        self.m_generate_call = self.m_generate_call.labels(name)
-        self.m_generate_success = self.m_generate_success.labels(name)
-        self.m_generate_failure = self.m_generate_failure.labels(name)
-        self.detokenizer = nltk.tokenize.treebank.TreebankWordDetokenizer()
-
-    @property
-    def tokenizer(self):
-        """
-        Return the tokenizer instance used by LLM
-        """
-        return self._get_tokenizer()
-
-    def _get_tokenizer(self):
-        pass
-
-    def has_structured_output(self):
-        # whether implementation supports structured output
-        # on the model side (otherwise it's prompt engineering)
-        return False
-
-    async def generate(
-        self,
-        prompt: str,
-        logger: reflector_logger,
-        gen_schema: dict | None = None,
-        gen_cfg: GenerationConfig | None = None,
-        **kwargs,
-    ) -> dict:
-        logger.info("LLM generate", prompt=repr(prompt))
-
-        if gen_cfg:
-            gen_cfg = gen_cfg.to_dict()
-        self.m_generate_call.inc()
-        try:
-            with self.m_generate.time():
-                result = await retry(self._generate)(
-                    prompt=prompt,
-                    gen_schema=gen_schema,
-                    gen_cfg=gen_cfg,
-                    logger=logger,
-                    **kwargs,
-                )
-            self.m_generate_success.inc()
-
-        except Exception:
-            logger.exception("Failed to call llm after retrying")
-            self.m_generate_failure.inc()
-            raise
-
-        logger.debug("LLM result [raw]", result=repr(result))
-        if isinstance(result, str):
-            result = self._parse_json(result)
-        logger.debug("LLM result [parsed]", result=repr(result))
-
-        return result
-
-    async def completion(
-        self, messages: list, logger: reflector_logger, **kwargs
-    ) -> dict:
-        """
-        Use /v1/chat/completion Open-AI compatible endpoint from the URL
-        It's up to the user to validate anything or transform the result
-        """
-        logger.info("LLM completions", messages=messages)
-
-        try:
-            with self.m_generate.time():
-                result = await retry(self._completion)(
-                    messages=messages, **{**kwargs, "logger": logger}
-                )
-            self.m_generate_success.inc()
-        except Exception:
-            logger.exception("Failed to call llm after retrying")
-            self.m_generate_failure.inc()
-            raise
-
-        logger.debug("LLM completion result", result=repr(result))
-        return result
-
-    def ensure_casing(self, title: str) -> str:
-        """
-        LLM takes care of word casing, but in rare cases this
-        can falter. This is a fallback to ensure the casing of
-        topics is in a proper format.
-
-        We select nouns, verbs and adjectives and check if camel
-         casing is present and fix it, if not. Will not perform
-         any other changes.
-        """
-        tokens = nltk.word_tokenize(title)
-        pos_tags = nltk.pos_tag(tokens)
-        camel_cased = []
-
-        whitelisted_pos_tags = [
-            "NN",
-            "NNS",
-            "NNP",
-            "NNPS",  # Noun POS
-            "VB",
-            "VBD",
-            "VBG",
-            "VBN",
-            "VBP",
-            "VBZ",  # Verb POS
-            "JJ",
-            "JJR",
-            "JJS",  # Adjective POS
-        ]
-
-        # If at all there is an exception, do not block other reflector
-        # processes. Return the LLM generated title, at the least.
-        try:
-            for word, pos in pos_tags:
-                if pos in whitelisted_pos_tags and word[0].islower():
-                    camel_cased.append(word[0].upper() + word[1:])
-                else:
-                    camel_cased.append(word)
-            modified_title = self.detokenizer.detokenize(camel_cased)
-
-            # Irrespective of casing changes, the starting letter
-            # of title is always upper-cased
-            title = modified_title[0].upper() + modified_title[1:]
-        except Exception as e:
-            reflector_logger.info(
-                f"Failed to ensure casing on {title=} with exception : {str(e)}"
-            )
-
-        return title
-
-    def trim_title(self, title: str) -> str:
-        """
-        List of manual trimming to the title.
-
-        Longer titles are prone to run into A prefix of phrases that don't
-        really add any descriptive information and in some cases, this
-        behaviour can be repeated for several consecutive topics. Trim the
-        titles to maintain quality of titles.
-        """
-        phrases_to_remove = ["Discussing", "Discussion on", "Discussion about"]
-        try:
-            pattern = (
-                r"\b(?:"
-                + "|".join(re.escape(phrase) for phrase in phrases_to_remove)
-                + r")\b"
-            )
-            title = re.sub(pattern, "", title, flags=re.IGNORECASE)
-        except Exception as e:
-            reflector_logger.info(f"Failed to trim {title=} with exception : {str(e)}")
-        return title
-
-    async def _generate(
-        self, prompt: str, gen_schema: dict | None, gen_cfg: dict | None, **kwargs
-    ) -> str:
-        raise NotImplementedError
-
-    async def _completion(self, messages: list, **kwargs) -> dict:
-        raise NotImplementedError
-
-    def _parse_json(self, result: str) -> dict:
-        result = result.strip()
-        # try detecting code block if exist
-        # starts with ```json\n, ends with ```
-        # or starts with ```\n, ends with ```
-        # or starts with \n```javascript\n, ends with ```
-
-        regex = r"```(json|javascript|)?(.*)```"
-        matches = re.findall(regex, result.strip(), re.MULTILINE | re.DOTALL)
-        if matches:
-            result = matches[0][1]
-
-        else:
-            # maybe the prompt has been started with ```json
-            # so if text ends with ```, just remove it and use it as json
-            if result.endswith("```"):
-                result = result[:-3]
-
-        return json.loads(result.strip())
-
-    def text_token_threshold(self, task_params: TaskParams | None) -> int:
-        """
-        Choose the token size to set as the threshold to pack the LLM calls
-        """
-        buffer_token_size = 100
-        default_output_tokens = 1000
-        context_window = self.tokenizer.model_max_length
-        tokens = self.tokenizer.tokenize(
-            self.create_prompt(instruct=task_params.instruct, text="")
-        )
-        threshold = context_window - len(tokens) - buffer_token_size
-        if task_params.gen_cfg:
-            threshold -= task_params.gen_cfg.max_new_tokens
-        else:
-            threshold -= default_output_tokens
-        return threshold
-
-    def split_corpus(
-        self,
-        corpus: str,
-        task_params: TaskParams,
-        token_threshold: int | None = None,
-    ) -> list[str]:
-        """
-        Split the input to the LLM due to CUDA memory limitations and LLM context window
-        restrictions.
-
-        Accumulate tokens from full sentences till threshold and yield accumulated
-        tokens. Reset accumulation when threshold is reached and repeat process.
-        """
-        if not token_threshold:
-            token_threshold = self.text_token_threshold(task_params=task_params)
-
-        accumulated_tokens = []
-        accumulated_sentences = []
-        accumulated_token_count = 0
-        corpus_sentences = nltk.sent_tokenize(corpus)
-
-        for sentence in corpus_sentences:
-            tokens = self.tokenizer.tokenize(sentence)
-            if accumulated_token_count + len(tokens) <= token_threshold:
-                accumulated_token_count += len(tokens)
-                accumulated_tokens.extend(tokens)
-                accumulated_sentences.append(sentence)
-            else:
-                yield "".join(accumulated_sentences)
-                accumulated_token_count = len(tokens)
-                accumulated_tokens = tokens
-                accumulated_sentences = [sentence]
-
-        if accumulated_tokens:
-            yield " ".join(accumulated_sentences)
-
-    def create_prompt(self, instruct: str, text: str) -> str:
-        """
-        Create a consumable prompt based on the prompt template
-        """
-        return self.template.format(instruct=instruct, text=text)

server/reflector/llm/llm_modal.py

@@ -1,155 +0,0 @@
-import httpx
-from transformers import AutoTokenizer, GenerationConfig
-
-from reflector.llm.base import LLM
-from reflector.logger import logger as reflector_logger
-from reflector.settings import settings
-from reflector.utils.retry import retry
-
-
-class ModalLLM(LLM):
-    def __init__(self, model_name: str | None = None):
-        super().__init__()
-        self.timeout = settings.LLM_TIMEOUT
-        self.llm_url = settings.LLM_URL + "/llm"
-        self.headers = {
-            "Authorization": f"Bearer {settings.LLM_MODAL_API_KEY}",
-        }
-        self._set_model_name(model_name if model_name else settings.DEFAULT_LLM)
-
-    @property
-    def supported_models(self):
-        """
-        List of currently supported models on this GPU platform
-        """
-        # TODO: Query the specific GPU platform
-        # Replace this with a HTTP call
-        return [
-            "lmsys/vicuna-13b-v1.5",
-            "HuggingFaceH4/zephyr-7b-alpha",
-            "NousResearch/Hermes-3-Llama-3.1-8B",
-        ]
-
-    async def _generate(
-        self, prompt: str, gen_schema: dict | None, gen_cfg: dict | None, **kwargs
-    ) -> str:
-        json_payload = {"prompt": prompt}
-        if gen_schema:
-            json_payload["gen_schema"] = gen_schema
-        if gen_cfg:
-            json_payload["gen_cfg"] = gen_cfg
-
-        # Handing over generation of the final summary to Zephyr model
-        # but replacing the Vicuna model will happen after more testing
-        # TODO: Create a mapping of model names and cloud deployments
-        if self.model_name == "HuggingFaceH4/zephyr-7b-alpha":
-            self.llm_url = settings.ZEPHYR_LLM_URL + "/llm"
-
-        async with httpx.AsyncClient() as client:
-            response = await retry(client.post)(
-                self.llm_url,
-                headers=self.headers,
-                json=json_payload,
-                timeout=self.timeout,
-                retry_timeout=60 * 5,
-                follow_redirects=True,
-                logger=kwargs.get("logger", reflector_logger),
-            )
-            response.raise_for_status()
-            text = response.json()["text"]
-            return text
-
-    async def _completion(self, messages: list, **kwargs) -> dict:
-        # returns full api response
-        kwargs.setdefault("temperature", 0.3)
-        kwargs.setdefault("max_tokens", 2048)
-        kwargs.setdefault("stream", False)
-        kwargs.setdefault("repetition_penalty", 1)
-        kwargs.setdefault("top_p", 1)
-        kwargs.setdefault("top_k", -1)
-        kwargs.setdefault("min_p", 0.05)
-        data = {"messages": messages, "model": self.model_name, **kwargs}
-
-        if self.model_name == "NousResearch/Hermes-3-Llama-3.1-8B":
-            self.llm_url = settings.HERMES_3_8B_LLM_URL + "/v1/chat/completions"
-
-        async with httpx.AsyncClient() as client:
-            response = await retry(client.post)(
-                self.llm_url,
-                headers=self.headers,
-                json=data,
-                timeout=self.timeout,
-                retry_timeout=60 * 5,
-                follow_redirects=True,
-                logger=kwargs.get("logger", reflector_logger),
-            )
-            response.raise_for_status()
-            return response.json()
-
-    def _set_model_name(self, model_name: str) -> bool:
-        """
-        Set the model name
-        """
-        # Abort, if the model is not supported
-        if model_name not in self.supported_models:
-            reflector_logger.info(
-                f"Attempted to change {model_name=}, but is not supported."
-                f"Setting model and tokenizer failed !"
-            )
-            return False
-        # Abort, if the model is already set
-        elif hasattr(self, "model_name") and model_name == self._get_model_name():
-            reflector_logger.info("No change in model. Setting model skipped.")
-            return False
-        # Update model name and tokenizer
-        self.model_name = model_name
-        self.llm_tokenizer = AutoTokenizer.from_pretrained(
-            self.model_name, cache_dir=settings.CACHE_DIR
-        )
-        reflector_logger.info(f"Model set to {model_name=}. Tokenizer updated.")
-        return True
-
-    def _get_tokenizer(self) -> AutoTokenizer:
-        """
-        Return the currently used LLM tokenizer
-        """
-        return self.llm_tokenizer
-
-    def _get_model_name(self) -> str:
-        """
-        Return the current model name from the instance details
-        """
-        return self.model_name
-
-
-LLM.register("modal", ModalLLM)
-
-if __name__ == "__main__":
-    from reflector.logger import logger
-
-    async def main():
-        llm = ModalLLM()
-        prompt = llm.create_prompt(
-            instruct="Complete the following task",
-            text="Tell me a joke about programming.",
-        )
-        result = await llm.generate(prompt=prompt, logger=logger)
-        print(result)
-
-        gen_schema = {
-            "type": "object",
-            "properties": {"response": {"type": "string"}},
-        }
-
-        result = await llm.generate(prompt=prompt, gen_schema=gen_schema, logger=logger)
-        print(result)
-
-        gen_cfg = GenerationConfig(max_new_tokens=150)
-        result = await llm.generate(
-            prompt=prompt, gen_cfg=gen_cfg, gen_schema=gen_schema, logger=logger
-        )
-        print(result)
-
-    import asyncio
-
-    asyncio.run(main())

server/reflector/llm/llm_openai.py

@@ -1,48 +0,0 @@
-import httpx
-from transformers import GenerationConfig
-
-from reflector.llm.base import LLM
-from reflector.logger import logger
-from reflector.settings import settings
-
-
-class OpenAILLM(LLM):
-    def __init__(self, model_name: str | None = None, **kwargs):
-        super().__init__(**kwargs)
-        self.openai_key = settings.LLM_OPENAI_KEY
-        self.openai_url = settings.LLM_URL
-        self.openai_model = settings.LLM_OPENAI_MODEL
-        self.openai_temperature = settings.LLM_OPENAI_TEMPERATURE
-        self.timeout = settings.LLM_TIMEOUT
-        self.max_tokens = settings.LLM_MAX_TOKENS
-        logger.info(f"LLM use openai backend at {self.openai_url}")
-
-    async def _generate(
-        self,
-        prompt: str,
-        gen_schema: dict | None,
-        gen_cfg: GenerationConfig | None,
-        **kwargs,
-    ) -> str:
-        headers = {
-            "Content-Type": "application/json",
-            "Authorization": f"Bearer {self.openai_key}",
-        }
-
-        async with httpx.AsyncClient(timeout=self.timeout) as client:
-            response = await client.post(
-                self.openai_url,
-                headers=headers,
-                json={
-                    "model": self.openai_model,
-                    "prompt": prompt,
-                    "max_tokens": self.max_tokens,
-                    "temperature": self.openai_temperature,
-                },
-            )
-            response.raise_for_status()
-            result = response.json()
-            return result["choices"][0]["text"]
-
-
-LLM.register("openai", OpenAILLM)

server/reflector/llm/llm_params.py

@@ -1,219 +0,0 @@
-from typing import Optional, TypeVar
-
-from pydantic import BaseModel
-from transformers import GenerationConfig
-
-
-class TaskParams(BaseModel, arbitrary_types_allowed=True):
-    instruct: str
-    gen_cfg: Optional[GenerationConfig] = None
-    gen_schema: Optional[dict] = None
-
-
-T = TypeVar("T", bound="LLMTaskParams")
-
-
-class LLMTaskParams:
-    _registry = {}
-
-    @classmethod
-    def register(cls, task, klass) -> None:
-        cls._registry[task] = klass
-
-    @classmethod
-    def get_instance(cls, task: str) -> T:
-        return cls._registry[task]()
-
-    @property
-    def task_params(self) -> TaskParams | None:
-        """
-        Fetch the task related parameters
-        """
-        return self._get_task_params()
-
-    def _get_task_params(self) -> None:
-        pass
-
-
-class FinalLongSummaryParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=1000, num_beams=3, do_sample=True, temperature=0.3
-        )
-        self._instruct = """
-        Take the key ideas and takeaways from the text and create a short
-         summary. Be sure to keep the length of the response to a minimum.
-         Do not include trivial information in the summary.
-          """
-        self._schema = {
-            "type": "object",
-            "properties": {"long_summary": {"type": "string"}},
-        }
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=self._schema, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """gen_schema
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-class FinalShortSummaryParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=800, num_beams=3, do_sample=True, temperature=0.3
-        )
-        self._instruct = """
-        Take the key ideas and takeaways from the text and create a short
-         summary. Be sure to keep the length of the response to a minimum.
-         Do not include trivial information in the summary.
-          """
-        self._schema = {
-            "type": "object",
-            "properties": {"short_summary": {"type": "string"}},
-        }
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=self._schema, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-class FinalTitleParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=200, num_beams=5, do_sample=True, temperature=0.5
-        )
-        self._instruct = """
-            Combine the following individual titles into one single short title that
-            condenses the essence of all titles.
-        """
-        self._schema = {
-            "type": "object",
-            "properties": {"title": {"type": "string"}},
-        }
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=self._schema, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-class TopicParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=500, num_beams=6, do_sample=True, temperature=0.9
-        )
-        self._instruct = """
-                Create a JSON object as response.The JSON object must have 2 fields:
-                i) title and ii) summary.
-                For the title field, generate a very detailed and self-explanatory
-                 title for the given text. Let the title be as descriptive as possible.
-                For the summary field, summarize the given text in a maximum of
-                two sentences.
-            """
-        self._schema = {
-            "type": "object",
-            "properties": {
-                "title": {"type": "string"},
-                "summary": {"type": "string"},
-            },
-        }
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=self._schema, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-class BulletedSummaryParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=800,
-            num_beams=1,
-            do_sample=True,
-            temperature=0.2,
-            early_stopping=True,
-        )
-        self._instruct = """
-        Given a meeting transcript, extract the key things discussed in the
-         form of a list.
-
-        While generating the response, follow the constraints mentioned below.
-
-        Summary constraints:
-        i) Do not add new content, except to fix spelling or punctuation.
-        ii) Do not add any prefixes or numbering in the response.
-        iii) The summarization should be as information dense as possible.
-        iv) Do not add any additional sections like Note, Conclusion, etc. in
-        the response.
-
-        Response format:
-        i) The response should be in the form of a bulleted list.
-        ii) Iteratively merge all the relevant paragraphs together to keep the
-         number of paragraphs to a minimum.
-        iii) Remove any unfinished sentences from the final response.
-        iv) Do not include narrative or reporting clauses.
-        v) Use "*" as the bullet icon.
-    """
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=None, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """gen_schema
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-class MergedSummaryParams(LLMTaskParams):
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-        self._gen_cfg = GenerationConfig(
-            max_new_tokens=600,
-            num_beams=1,
-            do_sample=True,
-            temperature=0.2,
-            early_stopping=True,
-        )
-        self._instruct = """
-        Given the key points of a meeting, summarize the points to describe the
-         meeting in the form of paragraphs.
-        """
-        self._task_params = TaskParams(
-            instruct=self._instruct, gen_schema=None, gen_cfg=self._gen_cfg
-        )
-
-    def _get_task_params(self) -> TaskParams:
-        """gen_schema
-        Return the parameters associated with a specific LLM task
-        """
-        return self._task_params
-
-
-LLMTaskParams.register("topic", TopicParams)
-LLMTaskParams.register("final_title", FinalTitleParams)
-LLMTaskParams.register("final_short_summary", FinalShortSummaryParams)
-LLMTaskParams.register("final_long_summary", FinalLongSummaryParams)
-LLMTaskParams.register("bullet_summary", BulletedSummaryParams)
-LLMTaskParams.register("merged_summary", MergedSummaryParams)

server/reflector/llm/openai_llm.py

@@ -1,118 +0,0 @@
-import httpx
-from transformers import AutoTokenizer
-
-from reflector.logger import logger
-
-
-def apply_gen_config(payload: dict, gen_cfg) -> None:
-    """Apply generation config overrides to the payload."""
-    config_mapping = {
-        "temperature": "temperature",
-        "max_new_tokens": "max_tokens",
-        "max_tokens": "max_tokens",
-        "top_p": "top_p",
-        "frequency_penalty": "frequency_penalty",
-        "presence_penalty": "presence_penalty",
-    }
-
-    for cfg_attr, payload_key in config_mapping.items():
-        value = getattr(gen_cfg, cfg_attr, None)
-        if value is not None:
-            payload[payload_key] = value
-            if cfg_attr == "max_new_tokens":  # Handle max_new_tokens taking precedence
-                break
-
-
-class OpenAILLM:
-    def __init__(self, config_prefix: str, settings):
-        self.config_prefix = config_prefix
-        self.settings_obj = settings
-        self.model_name = getattr(settings, f"{config_prefix}_MODEL")
-        self.url = getattr(settings, f"{config_prefix}_LLM_URL")
-        self.api_key = getattr(settings, f"{config_prefix}_LLM_API_KEY")
-
-        timeout = getattr(settings, f"{config_prefix}_LLM_TIMEOUT", 300)
-        self.temperature = getattr(settings, f"{config_prefix}_LLM_TEMPERATURE", 0.7)
-        self.max_tokens = getattr(settings, f"{config_prefix}_LLM_MAX_TOKENS", 1024)
-        self.client = httpx.AsyncClient(timeout=timeout)
-
-        # Use a tokenizer that approximates OpenAI token counting
-        tokenizer_name = getattr(settings, f"{config_prefix}_TOKENIZER", "gpt2")
-        try:
-            self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-        except Exception:
-            logger.debug(
-                f"Failed to load tokenizer '{tokenizer_name}', falling back to default 'gpt2' tokenizer"
-            )
-            self.tokenizer = AutoTokenizer.from_pretrained("gpt2")
-
-    async def generate(
-        self, prompt: str, gen_schema=None, gen_cfg=None, logger=None
-    ) -> str:
-        if logger:
-            logger.debug(
-                "OpenAI LLM generate",
-                prompt=repr(prompt[:100] + "..." if len(prompt) > 100 else prompt),
-            )
-
-        messages = [{"role": "user", "content": prompt}]
-        result = await self.completion(
-            messages, gen_schema=gen_schema, gen_cfg=gen_cfg, logger=logger
-        )
-        return result["choices"][0]["message"]["content"]
-
-    async def completion(
-        self, messages: list, gen_schema=None, gen_cfg=None, logger=None, **kwargs
-    ) -> dict:
-        if logger:
-            logger.info("OpenAI LLM completion", messages_count=len(messages))
-
-        payload = {
-            "model": self.model_name,
-            "messages": messages,
-            "temperature": self.temperature,
-            "max_tokens": self.max_tokens,
-        }
-
-        # Apply generation config overrides
-        if gen_cfg:
-            apply_gen_config(payload, gen_cfg)
-
-        # Apply structured output schema
-        if gen_schema:
-            payload["response_format"] = {
-                "type": "json_schema",
-                "json_schema": {"name": "response", "schema": gen_schema},
-            }
-
-        headers = {
-            "Content-Type": "application/json",
-            "Authorization": f"Bearer {self.api_key}",
-        }
-
-        url = f"{self.url.rstrip('/')}/chat/completions"
-
-        if logger:
-            logger.debug(
-                "OpenAI API request", url=url, payload_keys=list(payload.keys())
-            )
-
-        response = await self.client.post(url, json=payload, headers=headers)
-        response.raise_for_status()
-
-        result = response.json()
-
-        if logger:
-            logger.debug(
-                "OpenAI API response",
-                status_code=response.status_code,
-                choices_count=len(result.get("choices", [])),
-            )
-
-        return result
-
-    async def __aenter__(self):
-        return self
-
-    async def __aexit__(self, exc_type, exc_val, exc_tb):
-        await self.client.aclose()

server/reflector/pipelines/main_file_pipeline.py

@@ -0,0 +1,452 @@
+"""
+File-based processing pipeline
+==============================
+
+Optimized pipeline for processing complete audio/video files.
+Uses parallel processing for transcription, diarization, and waveform generation.
+"""
+
+import asyncio
+import uuid
+from pathlib import Path
+
+import av
+import structlog
+from celery import chain, shared_task
+from sqlalchemy.ext.asyncio import AsyncSession
+
+from reflector.asynctask import asynctask
+from reflector.db import get_session_factory
+from reflector.db.rooms import rooms_controller
+from reflector.db.transcripts import (
+    SourceKind,
+    Transcript,
+    TranscriptStatus,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines.main_live_pipeline import (
+    PipelineMainBase,
+    broadcast_to_sockets,
+    task_cleanup_consent,
+    task_pipeline_post_to_zulip,
+)
+from reflector.processors import (
+    AudioFileWriterProcessor,
+    TranscriptFinalSummaryProcessor,
+    TranscriptFinalTitleProcessor,
+    TranscriptTopicDetectorProcessor,
+)
+from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
+from reflector.processors.file_diarization import FileDiarizationInput
+from reflector.processors.file_diarization_auto import FileDiarizationAutoProcessor
+from reflector.processors.file_transcript import FileTranscriptInput
+from reflector.processors.file_transcript_auto import FileTranscriptAutoProcessor
+from reflector.processors.transcript_diarization_assembler import (
+    TranscriptDiarizationAssemblerInput,
+    TranscriptDiarizationAssemblerProcessor,
+)
+from reflector.processors.types import (
+    DiarizationSegment,
+    TitleSummary,
+)
+from reflector.processors.types import (
+    Transcript as TranscriptType,
+)
+from reflector.settings import settings
+from reflector.storage import get_transcripts_storage
+from reflector.worker.session_decorator import with_session
+from reflector.worker.webhook import send_transcript_webhook
+
+
+class EmptyPipeline:
+    """Empty pipeline for processors that need a pipeline reference"""
+
+    def __init__(self, logger: structlog.BoundLogger):
+        self.logger = logger
+
+    def get_pref(self, k, d=None):
+        return d
+
+    async def emit(self, event):
+        pass
+
+
+class PipelineMainFile(PipelineMainBase):
+    """
+    Optimized file processing pipeline.
+    Processes complete audio/video files with parallel execution.
+    """
+
+    logger: structlog.BoundLogger = None
+    empty_pipeline = None
+
+    def __init__(self, transcript_id: str):
+        super().__init__(transcript_id=transcript_id)
+        self.logger = logger.bind(transcript_id=self.transcript_id)
+        self.empty_pipeline = EmptyPipeline(logger=self.logger)
+
+    def _handle_gather_exceptions(self, results: list, operation: str) -> None:
+        """Handle exceptions from asyncio.gather with return_exceptions=True"""
+        for i, result in enumerate(results):
+            if not isinstance(result, Exception):
+                continue
+            self.logger.error(
+                f"Error in {operation} (task {i}): {result}",
+                transcript_id=self.transcript_id,
+                exc_info=result,
+            )
+
+    @broadcast_to_sockets
+    async def set_status(self, transcript_id: str, status: TranscriptStatus):
+        async with self.lock_transaction():
+            async with get_session_factory()() as session:
+                return await transcripts_controller.set_status(
+                    session, transcript_id, status
+                )
+
+    async def process(self, file_path: Path):
+        """Main entry point for file processing"""
+        self.logger.info(f"Starting file pipeline for {file_path}")
+
+        async with get_session_factory()() as session:
+            transcript = await transcripts_controller.get_by_id(
+                session, self.transcript_id
+            )
+
+            # Clear transcript as we're going to regenerate everything
+            await transcripts_controller.update(
+                session,
+                transcript,
+                {
+                    "events": [],
+                    "topics": [],
+                },
+            )
+
+        # Extract audio and write to transcript location
+        audio_path = await self.extract_and_write_audio(file_path, transcript)
+
+        # Upload for processing
+        audio_url = await self.upload_audio(audio_path, transcript)
+
+        # Run parallel processing
+        await self.run_parallel_processing(
+            session,
+            audio_path,
+            audio_url,
+            transcript.source_language,
+            transcript.target_language,
+        )
+
+        self.logger.info("File pipeline complete")
+
+        async with get_session_factory()() as session:
+            await transcripts_controller.set_status(session, transcript.id, "ended")
+
+    async def extract_and_write_audio(
+        self, file_path: Path, transcript: Transcript
+    ) -> Path:
+        """Extract audio from video if needed and write to transcript location as MP3"""
+        self.logger.info(f"Processing audio file: {file_path}")
+
+        # Check if it's already audio-only
+        container = av.open(str(file_path))
+        has_video = len(container.streams.video) > 0
+        container.close()
+
+        # Use AudioFileWriterProcessor to write MP3 to transcript location
+        mp3_writer = AudioFileWriterProcessor(
+            path=transcript.audio_mp3_filename,
+            on_duration=self.on_duration,
+        )
+
+        # Process audio frames and write to transcript location
+        input_container = av.open(str(file_path))
+        for frame in input_container.decode(audio=0):
+            await mp3_writer.push(frame)
+
+        await mp3_writer.flush()
+        input_container.close()
+
+        if has_video:
+            self.logger.info(
+                f"Extracted audio from video and saved to {transcript.audio_mp3_filename}"
+            )
+        else:
+            self.logger.info(
+                f"Converted audio file and saved to {transcript.audio_mp3_filename}"
+            )
+
+        return transcript.audio_mp3_filename
+
+    async def upload_audio(self, audio_path: Path, transcript: Transcript) -> str:
+        """Upload audio to storage for processing"""
+        storage = get_transcripts_storage()
+
+        if not storage:
+            raise Exception(
+                "Storage backend required for file processing. Configure TRANSCRIPT_STORAGE_* settings."
+            )
+
+        self.logger.info("Uploading audio to storage")
+
+        with open(audio_path, "rb") as f:
+            audio_data = f.read()
+
+        storage_path = f"file_pipeline/{transcript.id}/audio.mp3"
+        await storage.put_file(storage_path, audio_data)
+
+        audio_url = await storage.get_file_url(storage_path)
+
+        self.logger.info(f"Audio uploaded to {audio_url}")
+        return audio_url
+
+    async def run_parallel_processing(
+        self,
+        session,
+        audio_path: Path,
+        audio_url: str,
+        source_language: str,
+        target_language: str,
+    ):
+        """Coordinate parallel processing of transcription, diarization, and waveform"""
+        self.logger.info(
+            "Starting parallel processing", transcript_id=self.transcript_id
+        )
+
+        # Phase 1: Parallel processing of independent tasks
+        transcription_task = self.transcribe_file(audio_url, source_language)
+        diarization_task = self.diarize_file(audio_url)
+        waveform_task = self.generate_waveform(session, audio_path)
+
+        results = await asyncio.gather(
+            transcription_task, diarization_task, waveform_task, return_exceptions=True
+        )
+
+        transcript_result = results[0]
+        diarization_result = results[1]
+
+        # Handle errors - raise any exception that occurred
+        self._handle_gather_exceptions(results, "parallel processing")
+        for result in results:
+            if isinstance(result, Exception):
+                raise result
+
+        # Phase 2: Assemble transcript with diarization
+        self.logger.info(
+            "Assembling transcript with diarization", transcript_id=self.transcript_id
+        )
+        processor = TranscriptDiarizationAssemblerProcessor()
+        input_data = TranscriptDiarizationAssemblerInput(
+            transcript=transcript_result, diarization=diarization_result or []
+        )
+
+        # Store result for retrieval
+        diarized_transcript: Transcript | None = None
+
+        async def capture_result(transcript):
+            nonlocal diarized_transcript
+            diarized_transcript = transcript
+
+        processor.on(capture_result)
+        await processor.push(input_data)
+        await processor.flush()
+
+        if not diarized_transcript:
+            raise ValueError("No diarized transcript captured")
+
+        # Phase 3: Generate topics from diarized transcript
+        self.logger.info("Generating topics", transcript_id=self.transcript_id)
+        topics = await self.detect_topics(diarized_transcript, target_language)
+
+        # Phase 4: Generate title and summaries in parallel
+        self.logger.info(
+            "Generating title and summaries", transcript_id=self.transcript_id
+        )
+        results = await asyncio.gather(
+            self.generate_title(topics),
+            self.generate_summaries(session, topics),
+            return_exceptions=True,
+        )
+
+        self._handle_gather_exceptions(results, "title and summary generation")
+
+    async def transcribe_file(self, audio_url: str, language: str) -> TranscriptType:
+        """Transcribe complete file"""
+        processor = FileTranscriptAutoProcessor()
+        input_data = FileTranscriptInput(audio_url=audio_url, language=language)
+
+        # Store result for retrieval
+        result: TranscriptType | None = None
+
+        async def capture_result(transcript):
+            nonlocal result
+            result = transcript
+
+        processor.on(capture_result)
+        await processor.push(input_data)
+        await processor.flush()
+
+        if not result:
+            raise ValueError("No transcript captured")
+
+        return result
+
+    async def diarize_file(self, audio_url: str) -> list[DiarizationSegment] | None:
+        """Get diarization for file"""
+        if not settings.DIARIZATION_BACKEND:
+            self.logger.info("Diarization disabled")
+            return None
+
+        processor = FileDiarizationAutoProcessor()
+        input_data = FileDiarizationInput(audio_url=audio_url)
+
+        # Store result for retrieval
+        result = None
+
+        async def capture_result(diarization_output):
+            nonlocal result
+            result = diarization_output.diarization
+
+        try:
+            processor.on(capture_result)
+            await processor.push(input_data)
+            await processor.flush()
+            return result
+        except Exception as e:
+            self.logger.error(f"Diarization failed: {e}")
+            return None
+
+    async def generate_waveform(self, session: AsyncSession, audio_path: Path):
+        """Generate and save waveform"""
+        transcript = await transcripts_controller.get_by_id(session, self.transcript_id)
+
+        processor = AudioWaveformProcessor(
+            audio_path=audio_path,
+            waveform_path=transcript.audio_waveform_filename,
+            on_waveform=self.on_waveform,
+        )
+        processor.set_pipeline(self.empty_pipeline)
+
+        await processor.flush()
+
+    async def detect_topics(
+        self, transcript: TranscriptType, target_language: str
+    ) -> list[TitleSummary]:
+        """Detect topics from complete transcript"""
+        chunk_size = 300
+        topics: list[TitleSummary] = []
+
+        async def on_topic(topic: TitleSummary):
+            topics.append(topic)
+            return await self.on_topic(topic)
+
+        topic_detector = TranscriptTopicDetectorProcessor(callback=on_topic)
+        topic_detector.set_pipeline(self.empty_pipeline)
+
+        for i in range(0, len(transcript.words), chunk_size):
+            chunk_words = transcript.words[i : i + chunk_size]
+            if not chunk_words:
+                continue
+
+            chunk_transcript = TranscriptType(
+                words=chunk_words, translation=transcript.translation
+            )
+
+            await topic_detector.push(chunk_transcript)
+
+        await topic_detector.flush()
+        return topics
+
+    async def generate_title(self, topics: list[TitleSummary]):
+        """Generate title from topics"""
+        if not topics:
+            self.logger.warning("No topics for title generation")
+            return
+
+        processor = TranscriptFinalTitleProcessor(callback=self.on_title)
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+    async def generate_summaries(self, session, topics: list[TitleSummary]):
+        """Generate long and short summaries from topics"""
+        if not topics:
+            self.logger.warning("No topics for summary generation")
+            return
+
+        transcript = await transcripts_controller.get_by_id(session, self.transcript_id)
+        processor = TranscriptFinalSummaryProcessor(
+            transcript=transcript,
+            callback=self.on_long_summary,
+            on_short_summary=self.on_short_summary,
+        )
+        processor.set_pipeline(self.empty_pipeline)
+
+        for topic in topics:
+            await processor.push(topic)
+
+        await processor.flush()
+
+
+@shared_task
+@asynctask
+@with_session
+async def task_send_webhook_if_needed(session, *, transcript_id: str):
+    """Send webhook if this is a room recording with webhook configured"""
+    transcript = await transcripts_controller.get_by_id(session, transcript_id)
+    if not transcript:
+        return
+
+    if transcript.source_kind == SourceKind.ROOM and transcript.room_id:
+        room = await rooms_controller.get_by_id(session, transcript.room_id)
+        if room and room.webhook_url:
+            logger.info(
+                "Dispatching webhook",
+                transcript_id=transcript_id,
+                room_id=room.id,
+                webhook_url=room.webhook_url,
+            )
+            send_transcript_webhook.delay(
+                transcript_id, room.id, event_id=uuid.uuid4().hex
+            )
+
+
+@shared_task
+@asynctask
+@with_session
+async def task_pipeline_file_process(session, *, transcript_id: str):
+    """Celery task for file pipeline processing"""
+    transcript = await transcripts_controller.get_by_id(session, transcript_id)
+    if not transcript:
+        raise Exception(f"Transcript {transcript_id} not found")
+
+    pipeline = PipelineMainFile(transcript_id=transcript_id)
+    try:
+        await pipeline.set_status(transcript_id, "processing")
+
+        # Find the file to process
+        audio_file = next(transcript.data_path.glob("upload.*"), None)
+        if not audio_file:
+            audio_file = next(transcript.data_path.glob("audio.*"), None)
+
+        if not audio_file:
+            raise Exception("No audio file found to process")
+
+        await pipeline.process(audio_file)
+
+    except Exception:
+        await pipeline.set_status(transcript_id, "error")
+        raise
+
+    # Run post-processing chain: consent cleanup -> zulip -> webhook
+    post_chain = chain(
+        task_cleanup_consent.si(transcript_id=transcript_id),
+        task_pipeline_post_to_zulip.si(transcript_id=transcript_id),
+        task_send_webhook_if_needed.si(transcript_id=transcript_id),
+    )
+    post_chain.delay()

server/reflector/pipelines/main_live_pipeline.py

@@ -14,12 +14,17 @@ It is directly linked to our data model.
 import asyncio
 import functools
 from contextlib import asynccontextmanager
+from typing import Generic
 
+import av
 import boto3
 from celery import chord, current_task, group, shared_task
 from pydantic import BaseModel
+from sqlalchemy.ext.asyncio import AsyncSession
 from structlog import BoundLogger as Logger
 
+from reflector.asynctask import asynctask
+from reflector.db import get_session_factory
 from reflector.db.meetings import meeting_consent_controller, meetings_controller
 from reflector.db.recordings import recordings_controller
 from reflector.db.rooms import rooms_controller
@@ -29,16 +34,18 @@ from reflector.db.transcripts import (
     TranscriptFinalLongSummary,
     TranscriptFinalShortSummary,
     TranscriptFinalTitle,
+    TranscriptStatus,
     TranscriptText,
     TranscriptTopic,
     TranscriptWaveform,
     transcripts_controller,
 )
 from reflector.logger import logger
-from reflector.pipelines.runner import PipelineRunner
+from reflector.pipelines.runner import PipelineMessage, PipelineRunner
 from reflector.processors import (
-    AudioChunkerProcessor,
+    AudioChunkerAutoProcessor,
     AudioDiarizationAutoProcessor,
+    AudioDownscaleProcessor,
     AudioFileWriterProcessor,
     AudioMergeProcessor,
     AudioTranscriptAutoProcessor,
@@ -47,7 +54,7 @@ from reflector.processors import (
     TranscriptFinalTitleProcessor,
     TranscriptLinerProcessor,
     TranscriptTopicDetectorProcessor,
-    TranscriptTranslatorProcessor,
+    TranscriptTranslatorAutoProcessor,
 )
 from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
 from reflector.processors.types import AudioDiarizationInput
@@ -57,6 +64,7 @@ from reflector.processors.types import (
 from reflector.processors.types import Transcript as TranscriptProcessorType
 from reflector.settings import settings
 from reflector.storage import get_transcripts_storage
+from reflector.worker.session_decorator import with_session_and_transcript
 from reflector.ws_manager import WebsocketManager, get_ws_manager
 from reflector.zulip import (
     get_zulip_message,
@@ -65,30 +73,6 @@ from reflector.zulip import (
 )
 
 
-def asynctask(f):
-    @functools.wraps(f)
-    def wrapper(*args, **kwargs):
-        async def run_with_db():
-            from reflector.db import database
-
-            await database.connect()
-            try:
-                return await f(*args, **kwargs)
-            finally:
-                await database.disconnect()
-
-        coro = run_with_db()
-        try:
-            loop = asyncio.get_running_loop()
-        except RuntimeError:
-            loop = None
-        if loop and loop.is_running():
-            return loop.run_until_complete(coro)
-        return asyncio.run(coro)
-
-    return wrapper
-
-
 def broadcast_to_sockets(func):
     """
     Decorator to broadcast transcript event to websockets
@@ -115,9 +99,10 @@ def get_transcript(func):
     @functools.wraps(func)
     async def wrapper(**kwargs):
         transcript_id = kwargs.pop("transcript_id")
-        transcript = await transcripts_controller.get_by_id(transcript_id=transcript_id)
+        async with get_session_factory()() as session:
+            transcript = await transcripts_controller.get_by_id(session, transcript_id)
         if not transcript:
-            raise Exception("Transcript {transcript_id} not found")
+            raise Exception(f"Transcript {transcript_id} not found")
 
         # Enhanced logger with Celery task context
         tlogger = logger.bind(transcript_id=transcript.id)
@@ -144,27 +129,32 @@ class StrValue(BaseModel):
     value: str
 
 
-class PipelineMainBase(PipelineRunner):
-    transcript_id: str
-    ws_room_id: str | None = None
-    ws_manager: WebsocketManager | None = None
-
-    def prepare(self):
-        # prepare websocket
+class PipelineMainBase(PipelineRunner[PipelineMessage], Generic[PipelineMessage]):
+    def __init__(self, transcript_id: str):
+        super().__init__()
         self._lock = asyncio.Lock()
+        self.transcript_id = transcript_id
         self.ws_room_id = f"ts:{self.transcript_id}"
-        self.ws_manager = get_ws_manager()
+        self._ws_manager = None
 
-    async def get_transcript(self) -> Transcript:
+    @property
+    def ws_manager(self) -> WebsocketManager:
+        if self._ws_manager is None:
+            self._ws_manager = get_ws_manager()
+        return self._ws_manager
+
+    async def get_transcript(self, session: AsyncSession) -> Transcript:
         # fetch the transcript
-        result = await transcripts_controller.get_by_id(
-            transcript_id=self.transcript_id
-        )
+        result = await transcripts_controller.get_by_id(session, self.transcript_id)
         if not result:
             raise Exception("Transcript not found")
         return result
 
-    def get_transcript_topics(self, transcript: Transcript) -> list[TranscriptTopic]:
+    @staticmethod
+    def wrap_transcript_topics(
+        topics: list[TranscriptTopic],
+    ) -> list[TitleSummaryWithIdProcessorType]:
+        # transformation to a pipe-supported format
         return [
             TitleSummaryWithIdProcessorType(
                 id=topic.id,
@@ -174,28 +164,35 @@ class PipelineMainBase(PipelineRunner):
                 duration=topic.duration,
                 transcript=TranscriptProcessorType(words=topic.words),
             )
-            for topic in transcript.topics
+            for topic in topics
         ]
 
     @asynccontextmanager
-    async def transaction(self):
+    async def lock_transaction(self):
+        # This lock is to prevent multiple processor starting adding
+        # into event array at the same time
         async with self._lock:
-            async with transcripts_controller.transaction():
-                yield
+            yield
+
+    @asynccontextmanager
+    async def transaction(self):
+        async with self.lock_transaction():
+            async with get_session_factory()() as session:
+                yield session
 
     @broadcast_to_sockets
     async def on_status(self, status):
         # if it's the first part, update the status of the transcript
         # but do not set the ended status yet.
         if isinstance(self, PipelineMainLive):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "started": "recording",
                 "push": "recording",
                 "flush": "processing",
                 "error": "error",
             }
         elif isinstance(self, PipelineMainFinalSummaries):
-            status_mapping = {
+            status_mapping: dict[str, TranscriptStatus] = {
                 "push": "processing",
                 "flush": "processing",
                 "error": "error",
@@ -211,28 +208,18 @@ class PipelineMainBase(PipelineRunner):
             return
 
         # when the status of the pipeline changes, update the transcript
-        async with self.transaction():
-            transcript = await self.get_transcript()
-            if status == transcript.status:
-                return
-            resp = await transcripts_controller.append_event(
-                transcript=transcript,
-                event="STATUS",
-                data=StrValue(value=status),
-            )
-            await transcripts_controller.update(
-                transcript,
-                {
-                    "status": status,
-                },
-            )
-            return resp
+        async with self._lock:
+            async with get_session_factory()() as session:
+                return await transcripts_controller.set_status(
+                    session, self.transcript_id, status
+                )
 
     @broadcast_to_sockets
     async def on_transcript(self, data):
-        async with self.transaction():
-            transcript = await self.get_transcript()
+        async with self.transaction() as session:
+            transcript = await self.get_transcript(session)
             return await transcripts_controller.append_event(
+                session,
                 transcript=transcript,
                 event="TRANSCRIPT",
                 data=TranscriptText(text=data.text, translation=data.translation),
@@ -249,10 +236,11 @@ class PipelineMainBase(PipelineRunner):
         )
         if isinstance(data, TitleSummaryWithIdProcessorType):
             topic.id = data.id
-        async with self.transaction():
-            transcript = await self.get_transcript()
-            await transcripts_controller.upsert_topic(transcript, topic)
+        async with self.transaction() as session:
+            transcript = await self.get_transcript(session)
+            await transcripts_controller.upsert_topic(session, transcript, topic)
             return await transcripts_controller.append_event(
+                session,
                 transcript=transcript,
                 event="TOPIC",
                 data=topic,
@@ -261,16 +249,18 @@ class PipelineMainBase(PipelineRunner):
     @broadcast_to_sockets
     async def on_title(self, data):
         final_title = TranscriptFinalTitle(title=data.title)
-        async with self.transaction():
-            transcript = await self.get_transcript()
+        async with self.transaction() as session:
+            transcript = await self.get_transcript(session)
             if not transcript.title:
                 await transcripts_controller.update(
+                    session,
                     transcript,
                     {
                         "title": final_title.title,
                     },
                 )
             return await transcripts_controller.append_event(
+                session,
                 transcript=transcript,
                 event="FINAL_TITLE",
                 data=final_title,
@@ -279,15 +269,17 @@ class PipelineMainBase(PipelineRunner):
     @broadcast_to_sockets
     async def on_long_summary(self, data):
         final_long_summary = TranscriptFinalLongSummary(long_summary=data.long_summary)
-        async with self.transaction():
-            transcript = await self.get_transcript()
+        async with self.transaction() as session:
+            transcript = await self.get_transcript(session)
             await transcripts_controller.update(
+                session,
                 transcript,
                 {
                     "long_summary": final_long_summary.long_summary,
                 },
             )
             return await transcripts_controller.append_event(
+                session,
                 transcript=transcript,
                 event="FINAL_LONG_SUMMARY",
                 data=final_long_summary,
@@ -298,15 +290,17 @@ class PipelineMainBase(PipelineRunner):
         final_short_summary = TranscriptFinalShortSummary(
             short_summary=data.short_summary
         )
-        async with self.transaction():
-            transcript = await self.get_transcript()
+        async with self.transaction() as session:
+            transcript = await self.get_transcript(session)
             await transcripts_controller.update(
+                session,
                 transcript,
                 {
                     "short_summary": final_short_summary.short_summary,
                 },
             )
             return await transcripts_controller.append_event(
+                session,
                 transcript=transcript,
                 event="FINAL_SHORT_SUMMARY",
                 data=final_short_summary,
@@ -314,29 +308,30 @@ class PipelineMainBase(PipelineRunner):
 
     @broadcast_to_sockets
     async def on_duration(self, data):
-        async with self.transaction():
+        async with self.transaction() as session:
             duration = TranscriptDuration(duration=data)
 
-            transcript = await self.get_transcript()
+            transcript = await self.get_transcript(session)
             await transcripts_controller.update(
+                session,
                 transcript,
                 {
                     "duration": duration.duration,
                 },
             )
             return await transcripts_controller.append_event(
-                transcript=transcript, event="DURATION", data=duration
+                session, transcript=transcript, event="DURATION", data=duration
             )
 
     @broadcast_to_sockets
     async def on_waveform(self, data):
-        async with self.transaction():
+        async with self.transaction() as session:
             waveform = TranscriptWaveform(waveform=data)
 
-            transcript = await self.get_transcript()
+            transcript = await self.get_transcript(session)
 
             return await transcripts_controller.append_event(
-                transcript=transcript, event="WAVEFORM", data=waveform
+                session, transcript=transcript, event="WAVEFORM", data=waveform
             )
 
 
@@ -349,19 +344,20 @@ class PipelineMainLive(PipelineMainBase):
     async def create(self) -> Pipeline:
         # create a context for the whole rtc transaction
         # add a customised logger to the context
-        self.prepare()
-        transcript = await self.get_transcript()
+        async with get_session_factory()() as session:
+            transcript = await self.get_transcript(session)
 
         processors = [
             AudioFileWriterProcessor(
                 path=transcript.audio_wav_filename,
                 on_duration=self.on_duration,
             ),
-            AudioChunkerProcessor(),
+            AudioDownscaleProcessor(),
+            AudioChunkerAutoProcessor(),
             AudioMergeProcessor(),
             AudioTranscriptAutoProcessor.as_threaded(),
             TranscriptLinerProcessor(),
-            TranscriptTranslatorProcessor.as_threaded(callback=self.on_transcript),
+            TranscriptTranslatorAutoProcessor.as_threaded(callback=self.on_transcript),
             TranscriptTopicDetectorProcessor.as_threaded(callback=self.on_topic),
         ]
         pipeline = Pipeline(*processors)
@@ -370,6 +366,7 @@ class PipelineMainLive(PipelineMainBase):
         pipeline.set_pref("audio:target_language", transcript.target_language)
         pipeline.logger.bind(transcript_id=transcript.id)
         pipeline.logger.info("Pipeline main live created")
+        pipeline.describe()
 
         return pipeline
 
@@ -380,7 +377,7 @@ class PipelineMainLive(PipelineMainBase):
         pipeline_post(transcript_id=self.transcript_id)
 
 
-class PipelineMainDiarization(PipelineMainBase):
+class PipelineMainDiarization(PipelineMainBase[AudioDiarizationInput]):
     """
     Diarize the audio and update topics
     """
@@ -388,7 +385,6 @@ class PipelineMainDiarization(PipelineMainBase):
     async def create(self) -> Pipeline:
         # create a context for the whole rtc transaction
         # add a customised logger to the context
-        self.prepare()
         pipeline = Pipeline(
             AudioDiarizationAutoProcessor(callback=self.on_topic),
         )
@@ -397,18 +393,18 @@ class PipelineMainDiarization(PipelineMainBase):
         # now let's start the pipeline by pushing information to the
         # first processor diarization processor
         # XXX translation is lost when converting our data model to the processor model
-        transcript = await self.get_transcript()
+        async with get_session_factory()() as session:
+            transcript = await self.get_transcript(session)
 
         # diarization works only if the file is uploaded to an external storage
         if transcript.audio_location == "local":
             pipeline.logger.info("Audio is local, skipping diarization")
             return
 
-        topics = self.get_transcript_topics(transcript)
         audio_url = await transcript.get_audio_url()
         audio_diarization_input = AudioDiarizationInput(
             audio_url=audio_url,
-            topics=topics,
+            topics=self.wrap_transcript_topics(transcript.topics),
         )
 
         # as tempting to use pipeline.push, prefer to use the runner
@@ -421,7 +417,7 @@ class PipelineMainDiarization(PipelineMainBase):
         return pipeline
 
 
-class PipelineMainFromTopics(PipelineMainBase):
+class PipelineMainFromTopics(PipelineMainBase[TitleSummaryWithIdProcessorType]):
     """
     Pseudo class for generating a pipeline from topics
     """
@@ -430,10 +426,9 @@ class PipelineMainFromTopics(PipelineMainBase):
         raise NotImplementedError
 
     async def create(self) -> Pipeline:
-        self.prepare()
-
         # get transcript
-        self._transcript = transcript = await self.get_transcript()
+        async with get_session_factory()() as session:
+            self._transcript = transcript = await self.get_transcript(session)
 
         # create pipeline
         processors = self.get_processors()
@@ -443,7 +438,7 @@ class PipelineMainFromTopics(PipelineMainBase):
         pipeline.logger.info(f"{self.__class__.__name__} pipeline created")
 
         # push topics
-        topics = self.get_transcript_topics(transcript)
+        topics = PipelineMainBase.wrap_transcript_topics(transcript.topics)
         for topic in topics:
             await self.push(topic)
 
@@ -524,8 +519,6 @@ async def pipeline_convert_to_mp3(transcript: Transcript, logger: Logger):
     # Convert to mp3
     mp3_filename = transcript.audio_mp3_filename
 
-    import av
-
     with av.open(wav_filename.as_posix()) as in_container:
         in_stream = in_container.streams.audio[0]
         with av.open(mp3_filename.as_posix(), "w") as out_container:
@@ -540,8 +533,7 @@ async def pipeline_convert_to_mp3(transcript: Transcript, logger: Logger):
     logger.info("Convert to mp3 done")
 
 
-@get_transcript
-async def pipeline_upload_mp3(transcript: Transcript, logger: Logger):
+async def pipeline_upload_mp3(session, transcript: Transcript, logger: Logger):
     if not settings.TRANSCRIPT_STORAGE_BACKEND:
         logger.info("No storage backend configured, skipping mp3 upload")
         return
@@ -559,7 +551,7 @@ async def pipeline_upload_mp3(transcript: Transcript, logger: Logger):
         return
 
     # Upload to external storage and delete the file
-    await transcripts_controller.move_mp3_to_storage(transcript)
+    await transcripts_controller.move_mp3_to_storage(session, transcript)
 
     logger.info("Upload mp3 done")
 
@@ -588,23 +580,26 @@ async def pipeline_summaries(transcript: Transcript, logger: Logger):
     logger.info("Summaries done")
 
 
-@get_transcript
-async def cleanup_consent(transcript: Transcript, logger: Logger):
+async def cleanup_consent(session, transcript: Transcript, logger: Logger):
     logger.info("Starting consent cleanup")
 
     consent_denied = False
     recording = None
     try:
         if transcript.recording_id:
-            recording = await recordings_controller.get_by_id(transcript.recording_id)
+            recording = await recordings_controller.get_by_id(
+                session, transcript.recording_id
+            )
             if recording and recording.meeting_id:
-                meeting = await meetings_controller.get_by_id(recording.meeting_id)
+                meeting = await meetings_controller.get_by_id(
+                    session, recording.meeting_id
+                )
                 if meeting:
                     consent_denied = await meeting_consent_controller.has_any_denial(
-                        meeting.id
+                        session, meeting.id
                     )
     except Exception as e:
-        logger.error(f"Failed to get fetch consent: {e}")
+        logger.error(f"Failed to get fetch consent: {e}", exc_info=e)
         consent_denied = True
 
     if not consent_denied:
@@ -627,10 +622,10 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
                 f"Deleted original Whereby recording: {recording.bucket_name}/{recording.object_key}"
             )
         except Exception as e:
-            logger.error(f"Failed to delete Whereby recording: {e}")
+            logger.error(f"Failed to delete Whereby recording: {e}", exc_info=e)
 
     # non-transactional, files marked for deletion not actually deleted is possible
-    await transcripts_controller.update(transcript, {"audio_deleted": True})
+    await transcripts_controller.update(session, transcript, {"audio_deleted": True})
     # 2. Delete processed audio from transcript storage S3 bucket
     if transcript.audio_location == "storage":
         storage = get_transcripts_storage()
@@ -640,7 +635,7 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
                 f"Deleted processed audio from storage: {transcript.storage_audio_path}"
             )
         except Exception as e:
-            logger.error(f"Failed to delete processed audio: {e}")
+            logger.error(f"Failed to delete processed audio: {e}", exc_info=e)
 
     # 3. Delete local audio files
     try:
@@ -649,20 +644,19 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
         if hasattr(transcript, "audio_wav_filename") and transcript.audio_wav_filename:
             transcript.audio_wav_filename.unlink(missing_ok=True)
     except Exception as e:
-        logger.error(f"Failed to delete local audio files: {e}")
+        logger.error(f"Failed to delete local audio files: {e}", exc_info=e)
 
     logger.info("Consent cleanup done")
 
 
-@get_transcript
-async def pipeline_post_to_zulip(transcript: Transcript, logger: Logger):
+async def pipeline_post_to_zulip(session, transcript: Transcript, logger: Logger):
     logger.info("Starting post to zulip")
 
     if not transcript.recording_id:
         logger.info("Transcript has no recording")
         return
 
-    recording = await recordings_controller.get_by_id(transcript.recording_id)
+    recording = await recordings_controller.get_by_id(session, transcript.recording_id)
     if not recording:
         logger.info("Recording not found")
         return
@@ -671,12 +665,12 @@ async def pipeline_post_to_zulip(transcript: Transcript, logger: Logger):
         logger.info("Recording has no meeting")
         return
 
-    meeting = await meetings_controller.get_by_id(recording.meeting_id)
+    meeting = await meetings_controller.get_by_id(session, recording.meeting_id)
     if not meeting:
         logger.info("No meeting found for this recording")
         return
 
-    room = await rooms_controller.get_by_id(meeting.room_id)
+    room = await rooms_controller.get_by_id(session, meeting.room_id)
     if not room:
         logger.error(f"Missing room for a meeting {meeting.id}")
         return
@@ -702,7 +696,7 @@ async def pipeline_post_to_zulip(transcript: Transcript, logger: Logger):
                 room.zulip_stream, room.zulip_topic, message
             )
             await transcripts_controller.update(
-                transcript, {"zulip_message_id": response["id"]}
+                session, transcript, {"zulip_message_id": response["id"]}
             )
 
     logger.info("Posted to zulip")
@@ -733,8 +727,11 @@ async def task_pipeline_convert_to_mp3(*, transcript_id: str):
 
 @shared_task
 @asynctask
-async def task_pipeline_upload_mp3(*, transcript_id: str):
-    await pipeline_upload_mp3(transcript_id=transcript_id)
+@with_session_and_transcript
+async def task_pipeline_upload_mp3(
+    session, *, transcript: Transcript, logger: Logger, transcript_id: str
+):
+    await pipeline_upload_mp3(session, transcript=transcript, logger=logger)
 
 
 @shared_task
@@ -757,14 +754,20 @@ async def task_pipeline_final_summaries(*, transcript_id: str):
 
 @shared_task
 @asynctask
-async def task_cleanup_consent(*, transcript_id: str):
-    await cleanup_consent(transcript_id=transcript_id)
+@with_session_and_transcript
+async def task_cleanup_consent(
+    session, *, transcript: Transcript, logger: Logger, transcript_id: str
+):
+    await cleanup_consent(session, transcript=transcript, logger=logger)
 
 
 @shared_task
 @asynctask
-async def task_pipeline_post_to_zulip(*, transcript_id: str):
-    await pipeline_post_to_zulip(transcript_id=transcript_id)
+@with_session_and_transcript
+async def task_pipeline_post_to_zulip(
+    session, *, transcript: Transcript, logger: Logger, transcript_id: str
+):
+    await pipeline_post_to_zulip(session, transcript=transcript, logger=logger)
 
 
 def pipeline_post(*, transcript_id: str):
@@ -789,23 +792,23 @@ def pipeline_post(*, transcript_id: str):
         chain_final_summaries,
     ) | task_pipeline_post_to_zulip.si(transcript_id=transcript_id)
 
-    chain.delay()
+    return chain.delay()
 
 
 @get_transcript
 async def pipeline_process(transcript: Transcript, logger: Logger):
-    import av
-
     try:
         if transcript.audio_location == "storage":
-            await transcripts_controller.download_mp3_from_storage(transcript)
-            transcript.audio_waveform_filename.unlink(missing_ok=True)
-            await transcripts_controller.update(
-                transcript,
-                {
-                    "topics": [],
-                },
-            )
+            async with get_session_factory()() as session:
+                await transcripts_controller.download_mp3_from_storage(transcript)
+                transcript.audio_waveform_filename.unlink(missing_ok=True)
+                await transcripts_controller.update(
+                    session,
+                    transcript,
+                    {
+                        "topics": [],
+                    },
+                )
 
         # open audio
         audio_filename = next(transcript.data_path.glob("upload.*"), None)
@@ -837,12 +840,14 @@ async def pipeline_process(transcript: Transcript, logger: Logger):
 
     except Exception as exc:
         logger.error("Pipeline error", exc_info=exc)
-        await transcripts_controller.update(
-            transcript,
-            {
-                "status": "error",
-            },
-        )
+        async with get_session_factory()() as session:
+            await transcripts_controller.update(
+                session,
+                transcript,
+                {
+                    "status": "error",
+                },
+            )
         raise
 
     logger.info("Pipeline ended")

server/reflector/pipelines/runner.py

@@ -16,21 +16,16 @@ During its lifecycle, it will emit the following status:
 """
 
 import asyncio
-
-from pydantic import BaseModel, ConfigDict
+from typing import Generic, TypeVar
 
 from reflector.logger import logger
 from reflector.processors import Pipeline
 
+PipelineMessage = TypeVar("PipelineMessage")
 
-class PipelineRunner(BaseModel):
-    model_config = ConfigDict(arbitrary_types_allowed=True)
 
-    status: str = "idle"
-    pipeline: Pipeline | None = None
-
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
+class PipelineRunner(Generic[PipelineMessage]):
+    def __init__(self):
         self._task = None
         self._q_cmd = asyncio.Queue(maxsize=4096)
         self._ev_done = asyncio.Event()
@@ -39,6 +34,8 @@ class PipelineRunner(BaseModel):
             runner=id(self),
             runner_cls=self.__class__.__name__,
         )
+        self.status = "idle"
+        self.pipeline: Pipeline | None = None
 
     async def create(self) -> Pipeline:
         """
@@ -67,7 +64,7 @@ class PipelineRunner(BaseModel):
         coro = self.run()
         asyncio.run(coro)
 
-    async def push(self, data):
+    async def push(self, data: PipelineMessage):
         """
         Push data to the pipeline
         """
@@ -92,7 +89,11 @@ class PipelineRunner(BaseModel):
         pass
 
     async def _add_cmd(
-        self, cmd: str, data, max_retries: int = 3, retry_time_limit: int = 3
+        self,
+        cmd: str,
+        data: PipelineMessage,
+        max_retries: int = 3,
+        retry_time_limit: int = 3,
     ):
         """
         Enqueue a command to be executed in the runner.
@@ -143,7 +144,10 @@ class PipelineRunner(BaseModel):
                 cmd, data = await self._q_cmd.get()
                 func = getattr(self, f"cmd_{cmd.lower()}")
                 if func:
-                    await func(data)
+                    if cmd.upper() == "FLUSH":
+                        await func()
+                    else:
+                        await func(data)
                 else:
                     raise Exception(f"Unknown command {cmd}")
         except Exception:
@@ -152,13 +156,13 @@ class PipelineRunner(BaseModel):
             self._ev_done.set()
             raise
 
-    async def cmd_push(self, data):
+    async def cmd_push(self, data: PipelineMessage):
         if self._is_first_push:
             await self._set_status("push")
             self._is_first_push = False
         await self.pipeline.push(data)
 
-    async def cmd_flush(self, data):
+    async def cmd_flush(self):
         await self._set_status("flush")
         await self.pipeline.flush()
         await self._set_status("ended")

server/reflector/processors/__init__.py

@@ -1,5 +1,7 @@
 from .audio_chunker import AudioChunkerProcessor  # noqa: F401
+from .audio_chunker_auto import AudioChunkerAutoProcessor  # noqa: F401
 from .audio_diarization_auto import AudioDiarizationAutoProcessor  # noqa: F401
+from .audio_downscale import AudioDownscaleProcessor  # noqa: F401
 from .audio_file_writer import AudioFileWriterProcessor  # noqa: F401
 from .audio_merge import AudioMergeProcessor  # noqa: F401
 from .audio_transcript import AudioTranscriptProcessor  # noqa: F401
@@ -11,11 +13,19 @@ from .base import (  # noqa: F401
     Processor,
     ThreadedProcessor,
 )
+from .file_diarization import FileDiarizationProcessor  # noqa: F401
+from .file_diarization_auto import FileDiarizationAutoProcessor  # noqa: F401
+from .file_transcript import FileTranscriptProcessor  # noqa: F401
+from .file_transcript_auto import FileTranscriptAutoProcessor  # noqa: F401
+from .transcript_diarization_assembler import (
+    TranscriptDiarizationAssemblerProcessor,  # noqa: F401
+)
 from .transcript_final_summary import TranscriptFinalSummaryProcessor  # noqa: F401
 from .transcript_final_title import TranscriptFinalTitleProcessor  # noqa: F401
 from .transcript_liner import TranscriptLinerProcessor  # noqa: F401
 from .transcript_topic_detector import TranscriptTopicDetectorProcessor  # noqa: F401
 from .transcript_translator import TranscriptTranslatorProcessor  # noqa: F401
+from .transcript_translator_auto import TranscriptTranslatorAutoProcessor  # noqa: F401
 from .types import (  # noqa: F401
     AudioFile,
     FinalLongSummary,