fix: remove downscale from silero vad and use upstream processor

This commit restore the original behavior with frame cutting. While
silero is used on our gpu for files, look like it's not working great on
the live pipeline. To be investigated, but at the moment, what we keep
is:

- refactored to extract the downscale for further processing in the
pipeline
- remove any downscale implementation from audio_chunker and audio_merge
- removed batching from audio_merge too for now

.github/pull_request_template.md

@@ -1,19 +1,21 @@
-## ⚠️ Insert the PR TITLE replacing this text ⚠️
+<!--- Provide a general summary of your changes in the Title above -->
 
-⚠️ Describe your PR replacing this text. Post screenshots or videos whenever possible. ⚠️
+## Description
+<!--- Describe your changes in detail -->
 
-### Checklist
+## Related Issue
+<!--- This project only accepts pull requests related to open issues -->
+<!--- If suggesting a new feature or change, please discuss it in an issue first -->
+<!--- If fixing a bug, there should be an issue describing it with steps to reproduce -->
+<!--- Please link to the issue here: -->
 
- - [ ] My branch is updated with main (mandatory)
- - [ ] I wrote unit tests for this (if applies)
- - [ ] I have included migrations and tested them locally (if applies)
- - [ ] I have manually tested this feature locally
+## Motivation and Context
+<!--- Why is this change required? What problem does it solve? -->
+<!--- If it fixes an open issue, please link to the issue here. -->
 
-> IMPORTANT: Remember that you are responsible for merging this PR after it's been reviewed, and once deployed
-> you should perform manual testing to make sure everything went smoothly.
-
-### Urgency
-
- - [ ] Urgent (deploy ASAP)
- - [ ] Non-urgent (deploying in next release is ok)
+## How Has This Been Tested?
+<!--- Please describe in detail how you tested your changes. -->
+<!--- Include details of your testing environment, and the tests you ran to -->
+<!--- see how your change affects other areas of the code, etc. -->
 
+## Screenshots (if appropriate):

.github/workflows/conventional_commit_pr.yml

@@ -1,19 +0,0 @@
-name: Conventional commit PR
-
-on: [pull_request]
-
-jobs:
-  cog_check_job:
-    runs-on: ubuntu-latest
-    name: check conventional commit compliance
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          # pick the pr HEAD instead of the merge commit
-          ref: ${{ github.event.pull_request.head.sha }}
-
-      - name: Conventional commit check
-        uses: cocogitto/cocogitto-action@v3
-        with:
-          check-latest-tag-only: true

.github/workflows/db_migrations.yml

@@ -17,10 +17,40 @@ on:
 jobs:
   test-migrations:
     runs-on: ubuntu-latest
+    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_server.yml

@@ -19,29 +19,41 @@ 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
     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
+    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 +13,5 @@ restart-dev.sh
 data/
 www/REFACTOR.md
 www/reload-frontend
+server/test.sqlite
+CLAUDE.local.md

.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/
 
@@ -15,25 +15,15 @@ repos:
     hooks:
       - id: debug-statements
       - id: trailing-whitespace
-        exclude: ^server/trials
       - id: detect-private-key
 
-  - repo: https://github.com/psf/black
-    rev: 24.1.1
-    hooks:
-      - id: black
-        files: ^server/(reflector|tests)/
-
-  - repo: https://github.com/pycqa/isort
-    rev: 5.12.0
-    hooks:
-      - id: isort
-        name: isort (python)
-        files: ^server/(gpu|evaluate|reflector)/
-        args: [ "--profile", "black", "--filter-files" ]
-
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.6.5
+    rev: v0.8.2
     hooks:
       - id: ruff
-        files: ^server/(reflector|tests)/
+        args:
+          - --fix
+          # Uses select rules from server/pyproject.toml
+        files: ^server/
+      - id: ruff-format
+        files: ^server/

CHANGELOG.md

@@ -1,5 +1,76 @@
 # Changelog
 
+## [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)
+
+
+### Features
+
+* new summary using phi-4 and llama-index ([#519](https://github.com/Monadical-SAS/reflector/issues/519)) ([1bf9ce0](https://github.com/Monadical-SAS/reflector/commit/1bf9ce07c12f87f89e68a1dbb3b2c96c5ee62466))
+
+
+### Bug Fixes
+
+* remove unused settings and utils files ([#522](https://github.com/Monadical-SAS/reflector/issues/522)) ([2af4790](https://github.com/Monadical-SAS/reflector/commit/2af4790e4be9e588f282fbc1bb171c88a03d6479))
+
 ## [0.4.0](https://github.com/Monadical-SAS/reflector/compare/v0.3.2...v0.4.0) (2025-07-25)
 
 

CLAUDE.md

@@ -62,7 +62,7 @@ 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
@@ -72,19 +72,19 @@ 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 +144,9 @@ 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)
 
@@ -172,3 +174,7 @@ Modal.com integration for scalable ML processing:
 - **Audio Routing**: Use BlackHole (Mac) for merging multiple audio sources
 - **WebRTC**: Ensure proper CORS configuration for cross-origin streaming
 - **Database**: Run `uv run alembic upgrade head` after pulling schema changes
+
+## Pipeline/worker related info
+
+If you need to do any worker/pipeline related work, search for "Pipeline" classes and their "create" or "build" methods to find the main processor sequence. Look for task orchestration patterns (like "chord", "group", or "chain") to identify the post-processing flow with parallel execution chains. This will give you abstract vision on how processing pipeling is organized.

README.md

@@ -4,8 +4,8 @@
 
 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.
 
-[![Tests](https://github.com/monadical-sas/cubbi/actions/workflows/pytests.yml/badge.svg?branch=main&event=push)](https://github.com/monadical-sas/cubbi/actions/workflows/pytests.yml)
-[![License: MIT](https://img.shields.io/badge/license-AGPL--v3-green.svg)](https://opensource.org/licenses/AGPL-v3)
+[![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)
+[![License: MIT](https://img.shields.io/badge/license-MIT-green.svg)](https://opensource.org/licenses/MIT)
 </div>
 
 ## Screenshots
@@ -74,12 +74,12 @@ Note: We currently do not have instructions for Windows users.
 
 ### Frontend
 
-Start with `cd backend`.
+Start with `cd www`.
 
 **Installation**
 
 ```bash
-yarn install
+pnpm install
 cp .env_template .env
 cp config-template.ts config.ts
 ```
@@ -89,7 +89,7 @@ Then, fill in the environment variables in `.env` and the configuration in `conf
 **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 +99,7 @@ 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

compose.yml

@@ -39,11 +39,12 @@ 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
 

server/.env_template

@@ -1,16 +0,0 @@
-TRANSCRIPT_BACKEND=modal
-TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=***REMOVED***
-
-LLM_BACKEND=modal
-LLM_URL=https://monadical-sas--reflector-llm-web.modal.run
-LLM_MODAL_API_KEY=***REMOVED***
-
-TRANSLATE_URL=https://monadical-sas--reflector-translator-web.modal.run
-ZEPHYR_LLM_URL=https://monadical-sas--reflector-llm-zephyr-web.modal.run
-DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
-
-BASE_URL=https://xxxxx.ngrok.app
-DIARIZATION_ENABLED=false
-
-SQS_POLLING_TIMEOUT_SECONDS=60

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

@@ -20,3 +20,25 @@ Polls SQS every 60 seconds via /server/reflector/worker/process.py:24-62:
 # Every 60 seconds, check for new recordings
 sqs = boto3.client("sqs", ...)
 response = sqs.receive_message(QueueUrl=queue_url, ...)
+
+# Requeue
+
+```bash
+uv run /app/requeue_uploaded_file.py TRANSCRIPT_ID
+```
+
+## Pipeline Management
+
+### Continue stuck pipeline from final summaries (identify_participants) step:
+
+```bash
+uv run python -c "from reflector.pipelines.main_live_pipeline import task_pipeline_final_summaries; result = task_pipeline_final_summaries.delay(transcript_id='TRANSCRIPT_ID'); print(f'Task queued: {result.id}')"
+```
+
+### Run full post-processing pipeline (continues to completion):
+
+```bash
+uv run python -c "from reflector.pipelines.main_live_pipeline import pipeline_post; pipeline_post(transcript_id='TRANSCRIPT_ID')"
+```
+
+.

server/env.example

@@ -20,24 +20,24 @@ AUTH_JWT_AUDIENCE=
 
 ## Using local whisper
 #TRANSCRIPT_BACKEND=whisper
-#WHISPER_MODEL_SIZE=tiny
 
 ## 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
 TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-web.modal.run
-TRANSCRIPT_MODAL_API_KEY=***REMOVED***
+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
@@ -47,28 +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=***REMOVED***
-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
+## Context size for summary generation (tokens)
+# LLM_MODEL=microsoft/phi-4
+LLM_CONTEXT_WINDOW=16000
+LLM_URL=
+LLM_API_KEY=sk-
 
 ## =======================================================
 ## Diarization
@@ -77,7 +60,9 @@ CACHE_DIR=data
 ## 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

@@ -3,8 +3,10 @@
 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
+- `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
 
@@ -18,21 +20,29 @@ $ 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 theses keys:
+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
 
-LLM_BACKEND=modal
-LLM_URL=https://xxxx--reflector-llm-web.modal.run
-LLM_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
@@ -63,6 +73,86 @@ Authorization: bearer <REFLECTOR_APIKEY>
 
 ### 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)

server/gpu/modal_deployments/reflector_diarizer.py

@@ -4,14 +4,80 @@ Reflector GPU backend - diarizer
 """
 
 import os
+import uuid
+from typing import Mapping, NewType
+from urllib.parse import urlparse
 
-import modal.gpu
-from modal import App, Image, Secret, asgi_app, enter, method
-from pydantic import BaseModel
+import modal
 
 PYANNOTE_MODEL_NAME: str = "pyannote/speaker-diarization-3.1"
 MODEL_DIR = "/root/diarization_models"
-app = App(name="reflector-diarizer")
+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():
@@ -39,7 +105,7 @@ def download_pyannote_audio():
 
 
 diarizer_image = (
-    Image.debian_slim(python_version="3.10.8")
+    modal.Image.debian_slim(python_version="3.10.8")
     .pip_install(
         "pyannote.audio==3.1.0",
         "requests",
@@ -55,7 +121,8 @@ diarizer_image = (
         "hf-transfer",
     )
     .run_function(
-        download_pyannote_audio, secrets=[Secret.from_name("my-huggingface-secret")]
+        download_pyannote_audio,
+        secrets=[modal.Secret.from_name("hf_token")],
     )
     .run_function(migrate_cache_llm)
     .env(
@@ -70,53 +137,60 @@ diarizer_image = (
 
 
 @app.cls(
-    gpu=modal.gpu.A100(size="40GB"),
+    gpu="A100",
     timeout=60 * 30,
-    scaledown_window=60,
-    allow_concurrent_inputs=1,
     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:
-    @enter()
+    @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
+            PYANNOTE_MODEL_NAME,
+            cache_dir=MODEL_DIR,
+            use_auth_token=os.environ["HF_TOKEN"],
         )
         self.diarization_pipeline.to(torch.device(self.device))
 
-    @method()
-    def diarize(self, audio_data: str, audio_suffix: str, timestamp: float):
-        import tempfile
-
+    @modal.method()
+    def diarize(self, filename: str, timestamp: float = 0.0):
         import torchaudio
 
-        with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
-            fp.write(audio_data)
+        upload_volume.reload()
 
-            print("Diarizing audio")
-            waveform, sample_rate = torchaudio.load(fp.name)
-            diarization = self.diarization_pipeline(
-                {"waveform": waveform, "sample_rate": sample_rate}
+        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:]),
+                }
             )
-
-            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}
+        print("Diarization complete")
+        return {"diarization": words}
 
 
 # -------------------------------------------------------------------
@@ -127,17 +201,18 @@ class Diarizer:
 @app.function(
     timeout=60 * 10,
     scaledown_window=60 * 3,
-    allow_concurrent_inputs=40,
     secrets=[
-        Secret.from_name("reflector-gpu"),
+        modal.Secret.from_name("reflector-gpu"),
     ],
+    volumes={UPLOADS_PATH: upload_volume},
     image=diarizer_image,
 )
-@asgi_app()
+@modal.concurrent(max_inputs=40)
+@modal.asgi_app()
 def web():
-    import requests
     from fastapi import Depends, FastAPI, HTTPException, status
     from fastapi.security import OAuth2PasswordBearer
+    from pydantic import BaseModel
 
     diarizerstub = Diarizer()
 
@@ -153,35 +228,26 @@ def web():
                 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"
+    @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)
 
-        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
+        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

server/gpu/modal_deployments/reflector_llm.py

@@ -1,214 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-import modal
-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

@@ -1,220 +0,0 @@
-"""
-Reflector GPU backend - LLM
-===========================
-
-"""
-
-import json
-import os
-import threading
-from typing import Optional
-
-import modal
-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_parakeet.py

@@ -0,0 +1,622 @@
+import logging
+import os
+import sys
+import threading
+import uuid
+from typing import Mapping, NewType
+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 = {
+    "max_segment_duration": 30.0,
+    "batch_max_files": 10,
+    "batch_max_duration": 5.0,
+    "min_segment_duration": 0.02,
+    "silence_padding": 0.5,
+    "window_size": 512,
+}
+
+ParakeetUniqFilename = NewType("ParakeetUniqFilename", str)
+AudioFileExtension = NewType("AudioFileExtension", str)
+
+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 = [
+            {
+                "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 = [
+                {
+                    "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):
+            """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)
+
+                    # Extract the actual audio segment
+                    audio_segment = audio_array[start:end]
+
+                    yield (start_time, end_time, audio_segment)
+                    start = None
+
+            vad_iterator.reset_states()
+
+        def vad_segment_filter(segments):
+            """Filter VAD segments by duration and chunk large segments"""
+            min_dur = VAD_CONFIG["min_segment_duration"]
+            max_dur = VAD_CONFIG["max_segment_duration"]
+
+            for start_time, end_time, audio_segment in segments:
+                segment_duration = end_time - start_time
+
+                # Skip very small segments
+                if segment_duration < min_dur:
+                    continue
+
+                # If segment is within max duration, yield as-is
+                if segment_duration <= max_dur:
+                    yield (start_time, end_time, audio_segment)
+                    continue
+
+                # Chunk large segments into smaller pieces
+                chunk_samples = int(max_dur * SAMPLERATE)
+                current_start = start_time
+
+                for chunk_offset in range(0, len(audio_segment), chunk_samples):
+                    chunk_audio = audio_segment[
+                        chunk_offset : chunk_offset + chunk_samples
+                    ]
+                    if len(chunk_audio) == 0:
+                        break
+
+                    chunk_duration = len(chunk_audio) / float(SAMPLERATE)
+                    chunk_end = current_start + chunk_duration
+
+                    # Only yield chunks that meet minimum duration
+                    if chunk_duration >= min_dur:
+                        yield (current_start, chunk_end, chunk_audio)
+
+                    current_start = chunk_end
+
+        def batch_segments(segments, max_files=10, max_duration=5.0):
+            batch = []
+            batch_duration = 0.0
+
+            for start_time, end_time, audio_segment in segments:
+                segment_duration = end_time - start_time
+
+                if segment_duration < VAD_CONFIG["silence_padding"]:
+                    silence_samples = int(
+                        (VAD_CONFIG["silence_padding"] - segment_duration) * SAMPLERATE
+                    )
+                    padding = np.zeros(silence_samples, dtype=np.float32)
+                    audio_segment = np.concatenate([audio_segment, padding])
+                    segment_duration = VAD_CONFIG["silence_padding"]
+
+                batch.append((start_time, end_time, audio_segment))
+                batch_duration += segment_duration
+
+                if len(batch) >= max_files or batch_duration >= max_duration:
+                    yield batch
+                    batch = []
+                    batch_duration = 0.0
+
+            if batch:
+                yield batch
+
+        def transcribe_batch(model, audio_segments):
+            with NoStdStreams():
+                outputs = model.transcribe(audio_segments, timestamps=True)
+            return outputs
+
+        def emit_results(
+            results,
+            segments_info,
+            batch_index,
+            total_batches,
+        ):
+            """Yield transcribed text and word timings from model output, adjusting timestamps to absolute positions."""
+            for i, (output, (start_time, end_time, _)) in enumerate(
+                zip(results, segments_info)
+            ):
+                text = output.text.strip()
+                words = [
+                    {
+                        "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 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)
+        processed_duration = 0.0
+
+        all_text_parts = []
+        all_words = []
+
+        raw_segments = vad_segment_generator(audio_array)
+        filtered_segments = vad_segment_filter(raw_segments)
+        batches = batch_segments(
+            filtered_segments,
+            VAD_CONFIG["batch_max_files"],
+            VAD_CONFIG["batch_max_duration"],
+        )
+
+        batch_index = 0
+        total_batches = max(
+            1, int(total_duration / VAD_CONFIG["batch_max_duration"]) + 1
+        )
+
+        for batch in batches:
+            batch_index += 1
+            audio_segments = [seg[2] for seg in batch]
+            results = transcribe_batch(self.model, audio_segments)
+
+            for text, words in emit_results(
+                results,
+                batch,
+                batch_index,
+                total_batches,
+            ):
+                if not text:
+                    continue
+                all_text_parts.append(text)
+                all_words.extend(words)
+
+            processed_duration += sum(len(seg[2]) / float(SAMPLERATE) for seg in batch)
+
+        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()

server/gpu/modal_deployments/reflector_vllm_hermes3.py

@@ -1,171 +0,0 @@
-# # Run an OpenAI-Compatible vLLM Server
-
-import modal
-
-MODELS_DIR = "/llamas"
-MODEL_NAME = "NousResearch/Hermes-3-Llama-3.1-8B"
-N_GPU = 1
-
-
-def download_llm():
-    from huggingface_hub import snapshot_download
-
-    print("Downloading LLM model")
-    snapshot_download(
-        MODEL_NAME,
-        local_dir=f"{MODELS_DIR}/{MODEL_NAME}",
-        ignore_patterns=[
-            "*.pt",
-            "*.bin",
-            "*.pth",
-            "original/*",
-        ],  # Ensure safetensors
-    )
-    print("LLM model downloaded")
-
-
-def move_cache():
-    from transformers.utils import move_cache as transformers_move_cache
-
-    transformers_move_cache()
-
-
-vllm_image = (
-    modal.Image.debian_slim(python_version="3.10")
-    .pip_install("vllm==0.5.3post1")
-    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
-    .pip_install(
-        # "accelerate==0.34.2",
-        "einops==0.8.0",
-        "hf-transfer~=0.1",
-    )
-    .run_function(download_llm)
-    .run_function(move_cache)
-    .pip_install(
-        "bitsandbytes>=0.42.9",
-    )
-)
-
-app = modal.App("reflector-vllm-hermes3")
-
-
-@app.function(
-    image=vllm_image,
-    gpu=modal.gpu.A100(count=N_GPU, size="40GB"),
-    timeout=60 * 5,
-    scaledown_window=60 * 5,
-    allow_concurrent_inputs=100,
-    secrets=[
-        modal.Secret.from_name("reflector-gpu"),
-    ],
-)
-@modal.asgi_app()
-def serve():
-    import os
-
-    import fastapi
-    import vllm.entrypoints.openai.api_server as api_server
-    from vllm.engine.arg_utils import AsyncEngineArgs
-    from vllm.engine.async_llm_engine import AsyncLLMEngine
-    from vllm.entrypoints.logger import RequestLogger
-    from vllm.entrypoints.openai.serving_chat import OpenAIServingChat
-    from vllm.entrypoints.openai.serving_completion import OpenAIServingCompletion
-    from vllm.usage.usage_lib import UsageContext
-
-    TOKEN = os.environ["REFLECTOR_GPU_APIKEY"]
-
-    # create a fastAPI app that uses vLLM's OpenAI-compatible router
-    web_app = fastapi.FastAPI(
-        title=f"OpenAI-compatible {MODEL_NAME} server",
-        description="Run an OpenAI-compatible LLM server with vLLM on modal.com",
-        version="0.0.1",
-        docs_url="/docs",
-    )
-
-    # security: CORS middleware for external requests
-    http_bearer = fastapi.security.HTTPBearer(
-        scheme_name="Bearer Token",
-        description="See code for authentication details.",
-    )
-    web_app.add_middleware(
-        fastapi.middleware.cors.CORSMiddleware,
-        allow_origins=["*"],
-        allow_credentials=True,
-        allow_methods=["*"],
-        allow_headers=["*"],
-    )
-
-    # security: inject dependency on authed routes
-    async def is_authenticated(api_key: str = fastapi.Security(http_bearer)):
-        if api_key.credentials != TOKEN:
-            raise fastapi.HTTPException(
-                status_code=fastapi.status.HTTP_401_UNAUTHORIZED,
-                detail="Invalid authentication credentials",
-            )
-        return {"username": "authenticated_user"}
-
-    router = fastapi.APIRouter(dependencies=[fastapi.Depends(is_authenticated)])
-
-    # wrap vllm's router in auth router
-    router.include_router(api_server.router)
-    # add authed vllm to our fastAPI app
-    web_app.include_router(router)
-
-    engine_args = AsyncEngineArgs(
-        model=MODELS_DIR + "/" + MODEL_NAME,
-        tensor_parallel_size=N_GPU,
-        gpu_memory_utilization=0.90,
-        # max_model_len=8096,
-        enforce_eager=False,  # capture the graph for faster inference, but slower cold starts (30s > 20s)
-        # --- 4 bits load
-        # quantization="bitsandbytes",
-        # load_format="bitsandbytes",
-    )
-
-    engine = AsyncLLMEngine.from_engine_args(
-        engine_args, usage_context=UsageContext.OPENAI_API_SERVER
-    )
-
-    model_config = get_model_config(engine)
-
-    request_logger = RequestLogger(max_log_len=2048)
-
-    api_server.openai_serving_chat = OpenAIServingChat(
-        engine,
-        model_config=model_config,
-        served_model_names=[MODEL_NAME],
-        chat_template=None,
-        response_role="assistant",
-        lora_modules=[],
-        prompt_adapters=[],
-        request_logger=request_logger,
-    )
-    api_server.openai_serving_completion = OpenAIServingCompletion(
-        engine,
-        model_config=model_config,
-        served_model_names=[MODEL_NAME],
-        lora_modules=[],
-        prompt_adapters=[],
-        request_logger=request_logger,
-    )
-
-    return web_app
-
-
-def get_model_config(engine):
-    import asyncio
-
-    try:  # adapted from vLLM source -- https://github.com/vllm-project/vllm/blob/507ef787d85dec24490069ffceacbd6b161f4f72/vllm/entrypoints/openai/api_server.py#L235C1-L247C1
-        event_loop = asyncio.get_running_loop()
-    except RuntimeError:
-        event_loop = None
-
-    if event_loop is not None and event_loop.is_running():
-        # If the current is instanced by Ray Serve,
-        # there is already a running event loop
-        model_config = event_loop.run_until_complete(engine.get_model_config())
-    else:
-        # When using single vLLM without engine_use_ray
-        model_config = asyncio.run(engine.get_model_config())
-
-    return model_config

server/migration.load

@@ -1,16 +0,0 @@
-LOAD DATABASE
-     FROM sqlite:///app/reflector.sqlite3
-     INTO pgsql://reflector:reflector@postgres:5432/reflector
-WITH
-    include drop,
-    create tables,
-    create indexes,
-    reset sequences,
-    preserve index names,
-    prefetch rows = 10
-SET
-    work_mem to '512MB',
-    maintenance_work_mem to '1024MB'
-CAST
-    column transcript.duration to float using (lambda (val) (when val (format nil "~f" val)))
-;

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

@@ -1,9 +1,10 @@
 from logging.config import fileConfig
 
 from alembic import context
+from sqlalchemy import engine_from_config, pool
+
 from reflector.db import metadata
 from reflector.settings import settings
-from sqlalchemy import engine_from_config, pool
 
 # this is the Alembic Config object, which provides
 # access to the values within the .ini file in use.

server/migrations/versions/0925da921477_unique_room_names.py

@@ -8,7 +8,6 @@ Create Date: 2024-09-24 16:12:56.944133
 
 from typing import Sequence, Union
 
-import sqlalchemy as sa
 from alembic import op
 
 # revision identifiers, used by Alembic.

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/0fea6d96b096_add_share_mode.py

@@ -5,11 +5,11 @@ Revises: f819277e5169
 Create Date: 2023-11-07 11:12:21.614198
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "0fea6d96b096"

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/125031f7cb78_participants.py

@@ -5,26 +5,26 @@ Revises: 0fea6d96b096
 Create Date: 2023-11-30 15:56:03.341466
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '125031f7cb78'
-down_revision: Union[str, None] = '0fea6d96b096'
+revision: str = "125031f7cb78"
+down_revision: Union[str, None] = "0fea6d96b096"
 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.add_column('transcript', sa.Column('participants', sa.JSON(), nullable=True))
+    op.add_column("transcript", sa.Column("participants", sa.JSON(), nullable=True))
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_column('transcript', 'participants')
+    op.drop_column("transcript", "participants")
     # ### end Alembic commands ###

server/migrations/versions/20250617140003_add_meeting_consent_table.py

@@ -5,6 +5,7 @@ Revises: f819277e5169
 Create Date: 2025-06-17 14:00:03.000000
 
 """
+
 from typing import Sequence, Union
 
 import sqlalchemy as sa
@@ -19,16 +20,16 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     op.create_table(
-        'meeting_consent',
-        sa.Column('id', sa.String(), nullable=False),
-        sa.Column('meeting_id', sa.String(), nullable=False),
-        sa.Column('user_id', sa.String(), nullable=True),
-        sa.Column('consent_given', sa.Boolean(), nullable=False),
-        sa.Column('consent_timestamp', sa.DateTime(), nullable=False),
-        sa.PrimaryKeyConstraint('id'),
-        sa.ForeignKeyConstraint(['meeting_id'], ['meeting.id']),
+        "meeting_consent",
+        sa.Column("id", sa.String(), nullable=False),
+        sa.Column("meeting_id", sa.String(), nullable=False),
+        sa.Column("user_id", sa.String(), nullable=True),
+        sa.Column("consent_given", sa.Boolean(), nullable=False),
+        sa.Column("consent_timestamp", sa.DateTime(), nullable=False),
+        sa.PrimaryKeyConstraint("id"),
+        sa.ForeignKeyConstraint(["meeting_id"], ["meeting.id"]),
     )
 
 
 def downgrade() -> None:
-    op.drop_table('meeting_consent')
+    op.drop_table("meeting_consent")

server/migrations/versions/20250618140000_add_audio_deleted_field_to_transcript.py

@@ -5,6 +5,7 @@ Revises: 20250617140003
 Create Date: 2025-06-18 14:00:00.000000
 
 """
+
 from typing import Sequence, Union
 
 import sqlalchemy as sa
@@ -22,4 +23,4 @@ def upgrade() -> None:
 
 
 def downgrade() -> None:
-    op.drop_column("transcript", "audio_deleted")
+    op.drop_column("transcript", "audio_deleted")

server/migrations/versions/2cf0b60a9d34_fix_transcript_duration_type.py

@@ -5,36 +5,40 @@ Revises: ccd68dc784ff
 Create Date: 2025-07-15 16:53:40.397394
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '2cf0b60a9d34'
-down_revision: Union[str, None] = 'ccd68dc784ff'
+revision: str = "2cf0b60a9d34"
+down_revision: Union[str, None] = "ccd68dc784ff"
 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('transcript', schema=None) as batch_op:
-        batch_op.alter_column('duration',
-               existing_type=sa.INTEGER(),
-               type_=sa.Float(),
-               existing_nullable=True)
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "duration",
+            existing_type=sa.INTEGER(),
+            type_=sa.Float(),
+            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('duration',
-               existing_type=sa.Float(),
-               type_=sa.INTEGER(),
-               existing_nullable=True)
+    with op.batch_alter_table("transcript", schema=None) as batch_op:
+        batch_op.alter_column(
+            "duration",
+            existing_type=sa.Float(),
+            type_=sa.INTEGER(),
+            existing_nullable=True,
+        )
 
     # ### end Alembic commands ###

server/migrations/versions/38a927dcb099_rename_back_text_to_transcript.py

@@ -5,17 +5,17 @@ Revises: 9920ecfe2735
 Create Date: 2023-11-02 19:53:09.116240
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
-revision: str = '38a927dcb099'
-down_revision: Union[str, None] = '9920ecfe2735'
+revision: str = "38a927dcb099"
+down_revision: Union[str, None] = "9920ecfe2735"
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None
 

server/migrations/versions/4814901632bc_fix_duration.py

@@ -5,13 +5,13 @@ Revises: 38a927dcb099
 Create Date: 2023-11-10 18:12:17.886522
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
 revision: str = "4814901632bc"
@@ -24,9 +24,11 @@ def upgrade() -> None:
     # for all the transcripts, calculate the duration from the mp3
     # and update the duration column
     from pathlib import Path
-    from reflector.settings import settings
+
     import av
 
+    from reflector.settings import settings
+
     bind = op.get_bind()
     transcript = table(
         "transcript", column("id", sa.String), column("duration", sa.Float)

server/migrations/versions/543ed284d69a_init.py

@@ -5,14 +5,11 @@ Revises:
 Create Date: 2023-08-29 10:54:45.142974
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
-import sqlalchemy as sa
-
-
 # revision identifiers, used by Alembic.
-revision: str = '543ed284d69a'
+revision: str = "543ed284d69a"
 down_revision: Union[str, None] = None
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None

server/migrations/versions/62dea3db63a5_add_room_options.py

@@ -8,9 +8,8 @@ Create Date: 2025-06-27 09:04:21.006823
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "62dea3db63a5"
@@ -33,7 +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(
@@ -54,12 +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/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/764ce6db4388_add_zulip_message_id.py

@@ -5,26 +5,28 @@ Revises: 62dea3db63a5
 Create Date: 2024-09-06 14:02:06.649665
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = '764ce6db4388'
-down_revision: Union[str, None] = '62dea3db63a5'
+revision: str = "764ce6db4388"
+down_revision: Union[str, None] = "62dea3db63a5"
 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.add_column('transcript', sa.Column('zulip_message_id', sa.Integer(), nullable=True))
+    op.add_column(
+        "transcript", sa.Column("zulip_message_id", sa.Integer(), nullable=True)
+    )
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_column('transcript', 'zulip_message_id')
+    op.drop_column("transcript", "zulip_message_id")
     # ### 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/88d292678ba2_fix_transcript_json_nan_values.py

@@ -9,8 +9,6 @@ Create Date: 2025-07-15 19:30:19.876332
 from typing import Sequence, Union
 
 from alembic import op
-import sqlalchemy as sa
-
 
 # revision identifiers, used by Alembic.
 revision: str = "88d292678ba2"
@@ -21,7 +19,7 @@ depends_on: Union[str, Sequence[str], None] = None
 
 def upgrade() -> None:
     import json
-    import re
+
     from sqlalchemy import text
 
     # Get database connection
@@ -58,7 +56,9 @@ def upgrade() -> None:
             fixed_events = json.dumps(jevents)
             assert "NaN" not in fixed_events
         except (json.JSONDecodeError, AssertionError) as e:
-            print(f"Warning: Invalid JSON for transcript {transcript_id}, skipping: {e}")
+            print(
+                f"Warning: Invalid JSON for transcript {transcript_id}, skipping: {e}"
+            )
             continue
 
         # Update the record with fixed JSON

server/migrations/versions/9920ecfe2735_rename_transcript_to_text.py

@@ -5,13 +5,13 @@ Revises: 99365b0cd87b
 Create Date: 2023-11-02 18:55:17.019498
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-from sqlalchemy.sql import table, column
+from alembic import op
 from sqlalchemy import select
-
+from sqlalchemy.sql import column, table
 
 # revision identifiers, used by Alembic.
 revision: str = "9920ecfe2735"

server/migrations/versions/99365b0cd87b_add_title_short_and_long_summary_and_.py

@@ -8,8 +8,8 @@ Create Date: 2023-09-01 20:19:47.216334
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "99365b0cd87b"
@@ -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/a9c9c229ee36_transcript_composite_index.py

@@ -9,8 +9,6 @@ Create Date: 2025-07-15 20:09:40.253018
 from typing import Sequence, Union
 
 from alembic import op
-import sqlalchemy as sa
-from sqlalchemy.dialects import postgresql
 
 # revision identifiers, used by Alembic.
 revision: str = "a9c9c229ee36"

server/migrations/versions/b0e5f7876032_add_meeting_is_active.py

@@ -5,30 +5,37 @@ Revises: 6ea59639f30e
 Create Date: 2025-01-28 10:06:50.446233
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = 'b0e5f7876032'
-down_revision: Union[str, None] = '6ea59639f30e'
+revision: str = "b0e5f7876032"
+down_revision: Union[str, None] = "6ea59639f30e"
 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.add_column(sa.Column('is_active', sa.Boolean(), server_default=sa.text('1'), nullable=False))
+    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("true"),
+                nullable=False,
+            )
+        )
 
     # ### 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_column('is_active')
+    with op.batch_alter_table("meeting", schema=None) as batch_op:
+        batch_op.drop_column("is_active")
 
     # ### end Alembic commands ###

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/b3df9681cae9_add_transcript_table.py

@@ -8,9 +8,8 @@ Create Date: 2025-06-27 08:57:16.306940
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "b3df9681cae9"

server/migrations/versions/b469348df210_remove_viewer_room_url.py

@@ -8,9 +8,8 @@ Create Date: 2024-10-11 13:45:28.914902
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "b469348df210"

server/migrations/versions/b7df9609542c_add_unique_constraint_one_active_.py

@@ -5,15 +5,15 @@ Revises: d7fbb74b673b
 Create Date: 2025-07-25 16:27:06.959868
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = 'b7df9609542c'
-down_revision: Union[str, None] = 'd7fbb74b673b'
+revision: str = "b7df9609542c"
+down_revision: Union[str, None] = "d7fbb74b673b"
 branch_labels: Union[str, Sequence[str], None] = None
 depends_on: Union[str, Sequence[str], None] = None
 
@@ -22,14 +22,14 @@ def upgrade() -> None:
     # Create a partial unique index that ensures only one active meeting per room
     # This works for both PostgreSQL and SQLite
     op.create_index(
-        'idx_one_active_meeting_per_room',
-        'meeting',
-        ['room_id'],
+        "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')
+        postgresql_where=sa.text("is_active = true"),
+        sqlite_where=sa.text("is_active = 1"),
     )
 
 
 def downgrade() -> None:
-    op.drop_index('idx_one_active_meeting_per_room', table_name='meeting')
+    op.drop_index("idx_one_active_meeting_per_room", table_name="meeting")

server/migrations/versions/b9348748bbbc_reviewed.py

@@ -5,25 +5,31 @@ Revises: 125031f7cb78
 Create Date: 2023-12-13 15:37:51.303970
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
+from alembic import op
 
 # revision identifiers, used by Alembic.
-revision: str = 'b9348748bbbc'
-down_revision: Union[str, None] = '125031f7cb78'
+revision: str = "b9348748bbbc"
+down_revision: Union[str, None] = "125031f7cb78"
 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.add_column('transcript', sa.Column('reviewed', sa.Boolean(), server_default=sa.text('0'), nullable=False))
+    op.add_column(
+        "transcript",
+        sa.Column(
+            "reviewed", sa.Boolean(), server_default=sa.text("false"), nullable=False
+        ),
+    )
     # ### end Alembic commands ###
 
 
 def downgrade() -> None:
     # ### commands auto generated by Alembic - please adjust! ###
-    op.drop_column('transcript', 'reviewed')
+    op.drop_column("transcript", "reviewed")
     # ### end Alembic commands ###

server/migrations/versions/ccd68dc784ff_add_performance_indexes.py

@@ -9,8 +9,6 @@ Create Date: 2025-07-15 11:48:42.854741
 from typing import Sequence, Union
 
 from alembic import op
-import sqlalchemy as sa
-
 
 # revision identifiers, used by Alembic.
 revision: str = "ccd68dc784ff"

server/migrations/versions/d3ff3a39297f_add_recordings.py

@@ -8,9 +8,8 @@ Create Date: 2025-06-27 09:27:25.302152
 
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "d3ff3a39297f"

server/migrations/versions/d7fbb74b673b_add_room_id_to_transcript.py

@@ -56,4 +56,4 @@ def downgrade() -> None:
     op.drop_index("idx_transcript_room_id", "transcript")
 
     # Drop the room_id column
-    op.drop_column("transcript", "room_id")
+    op.drop_column("transcript", "room_id")

server/migrations/versions/f819277e5169_audio_location.py

@@ -5,11 +5,11 @@ Revises: 4814901632bc
 Create Date: 2023-11-16 10:29:09.351664
 
 """
+
 from typing import Sequence, Union
 
-from alembic import op
 import sqlalchemy as sa
-
+from alembic import op
 
 # revision identifiers, used by Alembic.
 revision: str = "f819277e5169"

server/pyproject.toml

@@ -32,12 +32,14 @@ dependencies = [
     "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",
 ]
 
 [dependency-groups]
@@ -54,6 +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 = [
@@ -62,6 +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 = [
@@ -69,6 +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]
@@ -81,10 +110,28 @@ packages = ["reflector"]
 [tool.coverage.run]
 source = ["reflector"]
 
+[tool.pytest_env]
+ENVIRONMENT = "pytest"
+DATABASE_URL = "postgresql://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"
+markers = [
+    "gpu_modal: mark test to run only with GPU Modal endpoints (deselect with '-m \"not gpu_modal\"')",
+]
+
+[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/**.py" = ["PLC0415"]
+"reflector/tools/**.py" = ["PLC0415"]
+"migrations/versions/**.py" = ["PLC0415"]
+"tests/**.py" = ["PLC0415"]

server/reflector/app.py

@@ -1,12 +1,13 @@
 from contextlib import asynccontextmanager
 
-import reflector.auth  # noqa
-import reflector.db  # noqa
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.routing import APIRoute
 from fastapi_pagination import add_pagination
 from prometheus_fastapi_instrumentator import Instrumentator
+
+import reflector.auth  # noqa
+import reflector.db  # noqa
 from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.logger import logger
 from reflector.metrics import metrics_init

server/reflector/auth/__init__.py

@@ -1,7 +1,8 @@
-from reflector.settings import settings
-from reflector.logger import logger
 import importlib
 
+from reflector.logger import logger
+from reflector.settings import settings
+
 logger.info(f"User authentication using {settings.AUTH_BACKEND}")
 module_name = f"reflector.auth.auth_{settings.AUTH_BACKEND}"
 auth_module = importlib.import_module(module_name)

server/reflector/auth/auth_jwt.py

@@ -4,6 +4,7 @@ from fastapi import Depends, HTTPException
 from fastapi.security import OAuth2PasswordBearer
 from jose import JWTError, jwt
 from pydantic import BaseModel
+
 from reflector.logger import logger
 from reflector.settings import settings
 

server/reflector/auth/auth_none.py

@@ -1,7 +1,8 @@
-from pydantic import BaseModel
 from typing import Annotated
+
 from fastapi import Depends
 from fastapi.security import OAuth2PasswordBearer
+from pydantic import BaseModel
 
 oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token", auto_error=False)
 

server/reflector/client.py

@@ -1,12 +1,12 @@
 import argparse
 import asyncio
 import signal
+from typing import NoReturn
 
 from aiortc.contrib.signaling import add_signaling_arguments, create_signaling
 
 from reflector.logger import logger
 from reflector.stream_client import StreamClient
-from typing import NoReturn
 
 
 async def main() -> NoReturn:
@@ -51,7 +51,7 @@ async def main() -> NoReturn:
 
         logger.info(f"Cancelling {len(tasks)} outstanding tasks")
         await asyncio.gather(*tasks, return_exceptions=True)
-        logger.info(f'{"Flushing metrics"}')
+        logger.info(f"{'Flushing metrics'}")
         loop.stop()
 
     signals = (signal.SIGHUP, signal.SIGTERM, signal.SIGINT)

server/reflector/db/__init__.py

@@ -1,11 +1,28 @@
+import contextvars
+from typing import Optional
+
 import databases
 import sqlalchemy
+
 from reflector.events import subscribers_shutdown, subscribers_startup
 from reflector.settings import settings
 
-database = databases.Database(settings.DATABASE_URL)
 metadata = sqlalchemy.MetaData()
 
+_database_context: contextvars.ContextVar[Optional[databases.Database]] = (
+    contextvars.ContextVar("database", default=None)
+)
+
+
+def get_database() -> databases.Database:
+    """Get database instance for current asyncio context"""
+    db = _database_context.get()
+    if db is None:
+        db = databases.Database(settings.DATABASE_URL)
+        _database_context.set(db)
+    return db
+
+
 # import models
 import reflector.db.meetings  # noqa
 import reflector.db.recordings  # noqa
@@ -13,16 +30,18 @@ import reflector.db.rooms  # noqa
 import reflector.db.transcripts  # noqa
 
 kwargs = {}
-if "sqlite" in settings.DATABASE_URL:
-    kwargs["connect_args"] = {"check_same_thread": False}
+if "postgres" not in settings.DATABASE_URL:
+    raise Exception("Only postgres database is supported in reflector")
 engine = sqlalchemy.create_engine(settings.DATABASE_URL, **kwargs)
 
 
 @subscribers_startup.append
 async def database_connect(_):
+    database = get_database()
     await database.connect()
 
 
 @subscribers_shutdown.append
 async def database_disconnect(_):
+    database = get_database()
     await database.disconnect()

server/reflector/db/meetings.py

@@ -4,7 +4,8 @@ from typing import Literal
 import sqlalchemy as sa
 from fastapi import HTTPException
 from pydantic import BaseModel, Field
-from reflector.db import database, metadata
+
+from reflector.db import get_database, metadata
 from reflector.db.rooms import Room
 from reflector.utils import generate_uuid4
 
@@ -15,8 +16,8 @@ meetings = sa.Table(
     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("start_date", sa.DateTime(timezone=True)),
+    sa.Column("end_date", sa.DateTime(timezone=True)),
     sa.Column("user_id", sa.String),
     sa.Column("room_id", sa.String),
     sa.Column("is_locked", sa.Boolean, nullable=False, server_default=sa.false()),
@@ -41,6 +42,12 @@ meetings = sa.Table(
         server_default=sa.true(),
     ),
     sa.Index("idx_meeting_room_id", "room_id"),
+    sa.Index(
+        "idx_one_active_meeting_per_room",
+        "room_id",
+        unique=True,
+        postgresql_where=sa.text("is_active = true"),
+    ),
 )
 
 meeting_consent = sa.Table(
@@ -50,7 +57,7 @@ meeting_consent = sa.Table(
     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),
+    sa.Column("consent_timestamp", sa.DateTime(timezone=True), nullable=False),
 )
 
 
@@ -110,7 +117,7 @@ class MeetingController:
             recording_trigger=room.recording_trigger,
         )
         query = meetings.insert().values(**meeting.model_dump())
-        await database.execute(query)
+        await get_database().execute(query)
         return meeting
 
     async def get_all_active(self) -> list[Meeting]:
@@ -118,7 +125,7 @@ class MeetingController:
         Get active meetings.
         """
         query = meetings.select().where(meetings.c.is_active)
-        return await database.fetch_all(query)
+        return await get_database().fetch_all(query)
 
     async def get_by_room_name(
         self,
@@ -128,7 +135,7 @@ class MeetingController:
         Get a meeting by room name.
         """
         query = meetings.select().where(meetings.c.room_name == room_name)
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
 
@@ -150,7 +157,7 @@ class MeetingController:
             )
             .order_by(end_date.desc())
         )
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
 
@@ -161,7 +168,7 @@ class MeetingController:
         Get a meeting by id
         """
         query = meetings.select().where(meetings.c.id == meeting_id)
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
         return Meeting(**result)
@@ -173,7 +180,7 @@ class MeetingController:
         If not found, it will raise a 404 error.
         """
         query = meetings.select().where(meetings.c.id == meeting_id)
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             raise HTTPException(status_code=404, detail="Meeting not found")
 
@@ -185,7 +192,7 @@ class MeetingController:
 
     async def update_meeting(self, meeting_id: str, **kwargs):
         query = meetings.update().where(meetings.c.id == meeting_id).values(**kwargs)
-        await database.execute(query)
+        await get_database().execute(query)
 
 
 class MeetingConsentController:
@@ -193,7 +200,7 @@ class MeetingConsentController:
         query = meeting_consent.select().where(
             meeting_consent.c.meeting_id == meeting_id
         )
-        results = await database.fetch_all(query)
+        results = await get_database().fetch_all(query)
         return [MeetingConsent(**result) for result in results]
 
     async def get_by_meeting_and_user(
@@ -204,7 +211,7 @@ class MeetingConsentController:
             meeting_consent.c.meeting_id == meeting_id,
             meeting_consent.c.user_id == user_id,
         )
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if result is None:
             return None
         return MeetingConsent(**result) if result else None
@@ -226,14 +233,14 @@ class MeetingConsentController:
                         consent_timestamp=consent.consent_timestamp,
                     )
                 )
-                await database.execute(query)
+                await get_database().execute(query)
 
                 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)
+        await get_database().execute(query)
         return consent
 
     async def has_any_denial(self, meeting_id: str) -> bool:
@@ -242,7 +249,7 @@ class MeetingConsentController:
             meeting_consent.c.meeting_id == meeting_id,
             meeting_consent.c.consent_given.is_(False),
         )
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         return result is not None
 
 

server/reflector/db/recordings.py

@@ -3,7 +3,8 @@ from typing import Literal
 
 import sqlalchemy as sa
 from pydantic import BaseModel, Field
-from reflector.db import database, metadata
+
+from reflector.db import get_database, metadata
 from reflector.utils import generate_uuid4
 
 recordings = sa.Table(
@@ -12,7 +13,7 @@ recordings = sa.Table(
     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("recorded_at", sa.DateTime(timezone=True), nullable=False),
     sa.Column(
         "status",
         sa.String,
@@ -36,12 +37,12 @@ class Recording(BaseModel):
 class RecordingController:
     async def create(self, recording: Recording):
         query = recordings.insert().values(**recording.model_dump())
-        await database.execute(query)
+        await get_database().execute(query)
         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)
+        result = await get_database().fetch_one(query)
         return Recording(**result) if result else None
 
     async def get_by_object_key(self, bucket_name: str, object_key: str) -> Recording:
@@ -49,8 +50,12 @@ class RecordingController:
             recordings.c.bucket_name == bucket_name,
             recordings.c.object_key == object_key,
         )
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         return Recording(**result) if result else None
 
+    async def remove_by_id(self, id: str) -> None:
+        query = recordings.delete().where(recordings.c.id == id)
+        await get_database().execute(query)
+
 
 recordings_controller = RecordingController()

server/reflector/db/rooms.py

@@ -1,21 +1,22 @@
-from datetime import datetime
+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 reflector.db import database, metadata
-from reflector.utils import generate_uuid4
 from sqlalchemy.sql import false, or_
 
+from reflector.db import get_database, metadata
+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("created_at", sqlalchemy.DateTime(timezone=True), nullable=False),
     sqlalchemy.Column(
         "zulip_auto_post", sqlalchemy.Boolean, nullable=False, server_default=false()
     ),
@@ -47,7 +48,7 @@ class Room(BaseModel):
     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 = ""
@@ -91,7 +92,7 @@ class RoomController:
         if return_query:
             return query
 
-        results = await database.fetch_all(query)
+        results = await get_database().fetch_all(query)
         return results
 
     async def add(
@@ -124,7 +125,7 @@ class RoomController:
         )
         query = rooms.insert().values(**room.model_dump())
         try:
-            await database.execute(query)
+            await get_database().execute(query)
         except IntegrityError:
             raise HTTPException(status_code=400, detail="Room name is not unique")
         return room
@@ -135,7 +136,7 @@ class RoomController:
         """
         query = rooms.update().where(rooms.c.id == room.id).values(**values)
         try:
-            await database.execute(query)
+            await get_database().execute(query)
         except IntegrityError:
             raise HTTPException(status_code=400, detail="Room name is not unique")
 
@@ -150,7 +151,7 @@ class RoomController:
         query = rooms.select().where(rooms.c.id == room_id)
         if "user_id" in kwargs:
             query = query.where(rooms.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
         return Room(**result)
@@ -162,7 +163,7 @@ class RoomController:
         query = rooms.select().where(rooms.c.name == room_name)
         if "user_id" in kwargs:
             query = query.where(rooms.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
         return Room(**result)
@@ -174,7 +175,7 @@ class RoomController:
         If not found, it will raise a 404 error.
         """
         query = rooms.select().where(rooms.c.id == meeting_id)
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             raise HTTPException(status_code=404, detail="Room not found")
 
@@ -196,7 +197,7 @@ class RoomController:
         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)
+        await get_database().execute(query)
 
 
 rooms_controller = RoomController()

server/reflector/db/search.py

@@ -0,0 +1,448 @@
+"""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,
+    ValidationError,
+    constr,
+    field_serializer,
+)
+
+from reflector.db import get_database
+from reflector.db.rooms import rooms
+from reflector.db.transcripts import SourceKind, transcripts
+from reflector.db.utils import is_postgresql
+from reflector.logger import logger
+
+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=0, 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")]
+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: str,
+        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
+    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: str = 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: str) -> 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
+        if not query:
+            logger.warning("Empty query for search text in count_matches")
+            return ZERO
+        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: str,
+        max_length: NonNegativeInt = DEFAULT_SNIPPET_MAX_LENGTH,
+        max_snippets: NonNegativeInt = DEFAULT_MAX_SNIPPETS,
+    ) -> list[str]:
+        """Generate snippets from text."""
+        if not text or not query:
+            logger.warning("Empty text or query 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: str,
+        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: str | None,
+        webvtt: WebVTTContent | None,
+        query: str,
+        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
+        """
+        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, params: SearchParameters
+    ) -> tuple[list[SearchResult], int]:
+        """
+        Full-text search for transcripts using PostgreSQL tsvector.
+        Returns (results, total_count).
+        """
+
+        if not is_postgresql():
+            logger.warning(
+                "Full-text search requires PostgreSQL. Returning empty results."
+            )
+            return [], 0
+
+        base_columns = [
+            transcripts.c.id,
+            transcripts.c.title,
+            transcripts.c.created_at,
+            transcripts.c.duration,
+            transcripts.c.status,
+            transcripts.c.user_id,
+            transcripts.c.room_id,
+            transcripts.c.source_kind,
+            transcripts.c.webvtt,
+            transcripts.c.long_summary,
+            sqlalchemy.case(
+                (
+                    transcripts.c.room_id.isnot(None) & rooms.c.id.is_(None),
+                    "Deleted Room",
+                ),
+                else_=rooms.c.name,
+            ).label("room_name"),
+        ]
+
+        if params.query_text:
+            search_query = sqlalchemy.func.websearch_to_tsquery(
+                "english", params.query_text
+            )
+            rank_column = sqlalchemy.func.ts_rank(
+                transcripts.c.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(
+            transcripts.join(rooms, transcripts.c.room_id == rooms.c.id, isouter=True)
+        )
+
+        if params.query_text:
+            base_query = base_query.where(
+                transcripts.c.search_vector_en.op("@@")(search_query)
+            )
+
+        if params.user_id:
+            base_query = base_query.where(
+                sqlalchemy.or_(
+                    transcripts.c.user_id == params.user_id, rooms.c.is_shared
+                )
+            )
+        else:
+            base_query = base_query.where(rooms.c.is_shared)
+        if params.room_id:
+            base_query = base_query.where(transcripts.c.room_id == params.room_id)
+        if params.source_kind:
+            base_query = base_query.where(
+                transcripts.c.source_kind == params.source_kind
+            )
+
+        if params.query_text:
+            order_by = sqlalchemy.desc(sqlalchemy.text("rank"))
+        else:
+            order_by = sqlalchemy.desc(transcripts.c.created_at)
+
+        query = base_query.order_by(order_by).limit(params.limit).offset(params.offset)
+
+        rs = await get_database().fetch_all(query)
+
+        count_query = sqlalchemy.select([sqlalchemy.func.count()]).select_from(
+            base_query.alias("search_results")
+        )
+        total = await get_database().fetch_val(count_query)
+
+        def _process_result(r) -> SearchResult:
+            r_dict: Dict[str, Any] = dict(r)
+            webvtt_raw: str | None = r_dict.pop("webvtt", None)
+            if webvtt_raw:
+                webvtt = WebVTTProcessor.parse(webvtt_raw)
+            else:
+                webvtt = None
+            long_summary: str | None = r_dict.pop("long_summary", None)
+            room_name: str | None = r_dict.pop("room_name", None)
+            db_result = SearchResultDB.model_validate(r_dict)
+
+            snippets, total_match_count = SnippetGenerator.combine_sources(
+                long_summary, webvtt, params.query_text, DEFAULT_MAX_SNIPPETS
+            )
+
+            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

@@ -3,20 +3,27 @@ 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 reflector.db import database, metadata
+from sqlalchemy import Enum
+from sqlalchemy.dialects.postgresql import TSVECTOR
+from sqlalchemy.sql import false, or_
+
+from reflector.db import get_database, metadata
+from reflector.db.recordings import recordings_controller
+from reflector.db.rooms import rooms
+from reflector.db.utils import is_postgresql
+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 sqlalchemy import Enum
-from sqlalchemy.sql import false, or_
+from reflector.utils.webvtt import topics_to_webvtt
 
 
 class SourceKind(enum.StrEnum):
@@ -33,7 +40,7 @@ transcripts = sqlalchemy.Table(
     sqlalchemy.Column("status", sqlalchemy.String),
     sqlalchemy.Column("locked", sqlalchemy.Boolean),
     sqlalchemy.Column("duration", sqlalchemy.Float),
-    sqlalchemy.Column("created_at", sqlalchemy.DateTime),
+    sqlalchemy.Column("created_at", sqlalchemy.DateTime(timezone=True)),
     sqlalchemy.Column("title", sqlalchemy.String),
     sqlalchemy.Column("short_summary", sqlalchemy.String),
     sqlalchemy.Column("long_summary", sqlalchemy.String),
@@ -75,13 +82,40 @@ transcripts = sqlalchemy.Table(
     # 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.Column("webvtt", sqlalchemy.Text),
     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"),
+    sqlalchemy.Index("idx_transcript_source_kind", "source_kind"),
+    sqlalchemy.Index("idx_transcript_room_id_created_at", "room_id", "created_at"),
 )
 
+# Add PostgreSQL-specific full-text search column
+# This matches the migration in migrations/versions/116b2f287eab_add_full_text_search.py
+if is_postgresql():
+    transcripts.append_column(
+        sqlalchemy.Column(
+            "search_vector_en",
+            TSVECTOR,
+            sqlalchemy.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,
+            ),
+        )
+    )
+    # Add GIN index for the search vector
+    transcripts.append_constraint(
+        sqlalchemy.Index(
+            "idx_transcript_search_vector_en",
+            "search_vector_en",
+            postgresql_using="gin",
+        )
+    )
+
 
 def generate_transcript_name() -> str:
     now = datetime.now(timezone.utc)
@@ -146,14 +180,18 @@ class TranscriptParticipant(BaseModel):
 
 
 class Transcript(BaseModel):
+    """Full transcript model with all fields."""
+
     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
     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] = []
@@ -167,9 +205,8 @@ 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:
@@ -270,10 +307,12 @@ 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",
@@ -334,7 +373,6 @@ 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
@@ -385,7 +423,7 @@ class TranscriptController:
         if return_query:
             return query
 
-        results = await database.fetch_all(query)
+        results = await get_database().fetch_all(query)
         return results
 
     async def get_by_id(self, transcript_id: str, **kwargs) -> Transcript | None:
@@ -395,7 +433,7 @@ class TranscriptController:
         query = transcripts.select().where(transcripts.c.id == transcript_id)
         if "user_id" in kwargs:
             query = query.where(transcripts.c.user_id == kwargs["user_id"])
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
         return Transcript(**result)
@@ -409,7 +447,7 @@ class TranscriptController:
         query = transcripts.select().where(transcripts.c.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)
+        result = await get_database().fetch_one(query)
         if not result:
             return None
         return Transcript(**result)
@@ -427,7 +465,7 @@ class TranscriptController:
             if order_by.startswith("-"):
                 field = field.desc()
             query = query.order_by(field)
-        results = await database.fetch_all(query)
+        results = await get_database().fetch_all(query)
         return [Transcript(**result) for result in results]
 
     async def get_by_id_for_http(
@@ -445,7 +483,7 @@ class TranscriptController:
         to determine if the user can access the transcript.
         """
         query = transcripts.select().where(transcripts.c.id == transcript_id)
-        result = await database.fetch_one(query)
+        result = await get_database().fetch_one(query)
         if not result:
             raise HTTPException(status_code=404, detail="Transcript not found")
 
@@ -498,23 +536,52 @@ class TranscriptController:
             room_id=room_id,
         )
         query = transcripts.insert().values(**transcript.model_dump())
-        await database.execute(query)
+        await get_database().execute(query)
         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, 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)
             .values(**values)
         )
-        await database.execute(query)
+        await get_database().execute(query)
         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,
         transcript_id: str,
@@ -528,23 +595,55 @@ class TranscriptController:
             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()
+        if transcript.recording_id:
+            try:
+                recording = await recordings_controller.get_by_id(
+                    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(transcript.recording_id)
+            except Exception as e:
+                logger.warning(
+                    "Failed to delete recording row",
+                    exc_info=e,
+                    recording_id=transcript.recording_id,
+                )
         query = transcripts.delete().where(transcripts.c.id == transcript_id)
-        await database.execute(query)
+        await get_database().execute(query)
 
     async def remove_by_recording_id(self, recording_id: str):
         """
         Remove a transcript by recording_id
         """
         query = transcripts.delete().where(transcripts.c.recording_id == recording_id)
-        await database.execute(query)
+        await get_database().execute(query)
 
     @asynccontextmanager
     async def transaction(self):
         """
         A context manager for database transaction
         """
-        async with database.transaction(isolation="serializable"):
+        async with get_database().transaction(isolation="serializable"):
             yield
 
     async def append_event(
@@ -557,11 +656,7 @@ 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(transcript, {"events": transcript.events_dump()})
         return resp
 
     async def upsert_topic(
@@ -573,11 +668,7 @@ class TranscriptController:
         Upsert topics to a transcript
         """
         transcript.upsert_topic(topic)
-        await self.update(
-            transcript,
-            {"topics": transcript.topics_dump()},
-            mutate=False,
-        )
+        await self.update(transcript, {"topics": transcript.topics_dump()})
 
     async def move_mp3_to_storage(self, transcript: Transcript):
         """
@@ -602,7 +693,8 @@ 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(transcript, {"audio_location": "storage"}, mutate=True)
 
         # unlink the local file
         transcript.audio_mp3_filename.unlink(missing_ok=True)
@@ -626,11 +718,7 @@ class TranscriptController:
         Add/update a participant to a transcript
         """
         result = transcript.upsert_participant(participant)
-        await self.update(
-            transcript,
-            {"participants": transcript.participants_dump()},
-            mutate=False,
-        )
+        await self.update(transcript, {"participants": transcript.participants_dump()})
         return result
 
     async def delete_participant(
@@ -642,11 +730,7 @@ class TranscriptController:
         Delete a participant from a transcript
         """
         transcript.delete_participant(participant_id)
-        await self.update(
-            transcript,
-            {"participants": transcript.participants_dump()},
-            mutate=False,
-        )
+        await self.update(transcript, {"participants": transcript.participants_dump()})
 
 
 transcripts_controller = TranscriptController()

server/reflector/db/utils.py

@@ -0,0 +1,9 @@
+"""Database utility functions."""
+
+from reflector.db import get_database
+
+
+def is_postgresql() -> bool:
+    return get_database().url.scheme and get_database().url.scheme.startswith(
+        "postgresql"
+    )

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

@@ -1,338 +0,0 @@
-import importlib
-import json
-import re
-from typing import TypeVar
-
-import nltk
-from prometheus_client import Counter, Histogram
-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
-from transformers import GenerationConfig
-
-T = TypeVar("T", bound="LLM")
-
-
-class LLM:
-    _nltk_downloaded = False
-    _registry = {}
-    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, defaults to `oobabooga`
-        - `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
-
-    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,
-                    **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)
-            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, logger: reflector_logger, **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

@@ -1,151 +0,0 @@
-import httpx
-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
-from transformers import AutoTokenizer, GenerationConfig
-
-
-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
-    ):
-        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,
-            )
-            response.raise_for_status()
-            text = response.json()["text"]
-            return text
-
-    async def _completion(self, messages: list, **kwargs) -> dict:
-        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,
-            )
-            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_oobabooga.py

@@ -1,29 +0,0 @@
-import httpx
-
-from reflector.llm.base import LLM
-from reflector.settings import settings
-
-
-class OobaboogaLLM(LLM):
-    def __init__(self, model_name: str | None = None):
-        super().__init__()
-
-    async def _generate(
-        self, prompt: str, gen_schema: dict | None, gen_cfg: dict | None, **kwargs
-    ):
-        json_payload = {"prompt": prompt}
-        if gen_schema:
-            json_payload["gen_schema"] = gen_schema
-        if gen_cfg:
-            json_payload.update(gen_cfg)
-        async with httpx.AsyncClient() as client:
-            response = await client.post(
-                settings.LLM_URL,
-                headers={"Content-Type": "application/json"},
-                json=json_payload,
-            )
-            response.raise_for_status()
-            return response.json()
-
-
-LLM.register("oobabooga", OobaboogaLLM)

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/pipelines/main_file_pipeline.py

@@ -0,0 +1,375 @@
+"""
+File-based processing pipeline
+==============================
+
+Optimized pipeline for processing complete audio/video files.
+Uses parallel processing for transcription, diarization, and waveform generation.
+"""
+
+import asyncio
+from pathlib import Path
+
+import av
+import structlog
+from celery import shared_task
+
+from reflector.db.transcripts import (
+    Transcript,
+    transcripts_controller,
+)
+from reflector.logger import logger
+from reflector.pipelines.main_live_pipeline import PipelineMainBase, asynctask
+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
+
+
+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,
+            )
+
+    async def process(self, file_path: Path):
+        """Main entry point for file processing"""
+        self.logger.info(f"Starting file pipeline for {file_path}")
+
+        transcript = await self.get_transcript()
+
+        # 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(
+            audio_path,
+            audio_url,
+            transcript.source_language,
+            transcript.target_language,
+        )
+
+        self.logger.info("File pipeline complete")
+
+    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,
+        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(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(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, audio_path: Path):
+        """Generate and save waveform"""
+        transcript = await self.get_transcript()
+
+        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, topics: list[TitleSummary]):
+        """Generate long and short summaries from topics"""
+        if not topics:
+            self.logger.warning("No topics for summary generation")
+            return
+
+        transcript = await self.get_transcript()
+        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
+async def task_pipeline_file_process(*, transcript_id: str):
+    """Celery task for file pipeline processing"""
+
+    transcript = await transcripts_controller.get_by_id(transcript_id)
+    if not transcript:
+        raise Exception(f"Transcript {transcript_id} not found")
+
+    # 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")
+
+    # Run file pipeline
+    pipeline = PipelineMainFile(transcript_id=transcript_id)
+    await pipeline.process(audio_file)

server/reflector/pipelines/main_live_pipeline.py

@@ -14,10 +14,15 @@ 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, group, shared_task
+from celery import chord, current_task, group, shared_task
 from pydantic import BaseModel
+from structlog import BoundLogger as Logger
+
+from reflector.db import get_database
 from reflector.db.meetings import meeting_consent_controller, meetings_controller
 from reflector.db.recordings import recordings_controller
 from reflector.db.rooms import rooms_controller
@@ -33,10 +38,11 @@ from reflector.db.transcripts import (
     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,
@@ -45,7 +51,7 @@ from reflector.processors import (
     TranscriptFinalTitleProcessor,
     TranscriptLinerProcessor,
     TranscriptTopicDetectorProcessor,
-    TranscriptTranslatorProcessor,
+    TranscriptTranslatorAutoProcessor,
 )
 from reflector.processors.audio_waveform_processor import AudioWaveformProcessor
 from reflector.processors.types import AudioDiarizationInput
@@ -61,15 +67,13 @@ from reflector.zulip import (
     send_message_to_zulip,
     update_zulip_message,
 )
-from structlog import BoundLogger as Logger
 
 
 def asynctask(f):
     @functools.wraps(f)
     def wrapper(*args, **kwargs):
         async def run_with_db():
-            from reflector.db import database
-
+            database = get_database()
             await database.connect()
             try:
                 return await f(*args, **kwargs)
@@ -111,16 +115,29 @@ def get_transcript(func):
     Decorator to fetch the transcript from the database from the first argument
     """
 
+    @functools.wraps(func)
     async def wrapper(**kwargs):
         transcript_id = kwargs.pop("transcript_id")
         transcript = await transcripts_controller.get_by_id(transcript_id=transcript_id)
         if not transcript:
             raise Exception("Transcript {transcript_id} not found")
+
+        # Enhanced logger with Celery task context
         tlogger = logger.bind(transcript_id=transcript.id)
+        if current_task:
+            tlogger = tlogger.bind(
+                task_id=current_task.request.id,
+                task_name=current_task.name,
+                worker_hostname=current_task.request.hostname,
+                task_retries=current_task.request.retries,
+                transcript_id=transcript_id,
+            )
+
         try:
-            return await func(transcript=transcript, logger=tlogger, **kwargs)
+            result = await func(transcript=transcript, logger=tlogger, **kwargs)
+            return result
         except Exception as exc:
-            tlogger.error("Pipeline error", exc_info=exc)
+            tlogger.error("Pipeline error", function_name=func.__name__, exc_info=exc)
             raise
 
     return wrapper
@@ -130,16 +147,19 @@ 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
+
+    @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) -> Transcript:
         # fetch the transcript
@@ -150,7 +170,11 @@ class PipelineMainBase(PipelineRunner):
             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,
@@ -160,7 +184,7 @@ class PipelineMainBase(PipelineRunner):
                 duration=topic.duration,
                 transcript=TranscriptProcessorType(words=topic.words),
             )
-            for topic in transcript.topics
+            for topic in topics
         ]
 
     @asynccontextmanager
@@ -335,7 +359,6 @@ 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()
 
         processors = [
@@ -343,11 +366,12 @@ class PipelineMainLive(PipelineMainBase):
                 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)
@@ -356,6 +380,7 @@ 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
 
@@ -366,7 +391,7 @@ class PipelineMainLive(PipelineMainBase):
         pipeline_post(transcript_id=self.transcript_id)
 
 
-class PipelineMainDiarization(PipelineMainBase):
+class PipelineMainDiarization(PipelineMainBase[AudioDiarizationInput]):
     """
     Diarize the audio and update topics
     """
@@ -374,7 +399,6 @@ 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),
         )
@@ -390,11 +414,10 @@ class PipelineMainDiarization(PipelineMainBase):
             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
@@ -407,7 +430,7 @@ class PipelineMainDiarization(PipelineMainBase):
         return pipeline
 
 
-class PipelineMainFromTopics(PipelineMainBase):
+class PipelineMainFromTopics(PipelineMainBase[TitleSummaryWithIdProcessorType]):
     """
     Pseudo class for generating a pipeline from topics
     """
@@ -416,8 +439,6 @@ class PipelineMainFromTopics(PipelineMainBase):
         raise NotImplementedError
 
     async def create(self) -> Pipeline:
-        self.prepare()
-
         # get transcript
         self._transcript = transcript = await self.get_transcript()
 
@@ -429,7 +450,7 @@ 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)
 
@@ -510,8 +531,6 @@ 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:
@@ -590,7 +609,7 @@ async def cleanup_consent(transcript: Transcript, logger: Logger):
                         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:
@@ -613,7 +632,7 @@ 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})
@@ -626,7 +645,7 @@ 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:
@@ -635,7 +654,7 @@ 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")
 
@@ -780,8 +799,6 @@ def pipeline_post(*, transcript_id: str):
 
 @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)

server/reflector/pipelines/runner.py

@@ -16,20 +16,16 @@ During its lifecycle, it will emit the following status:
 """
 
 import asyncio
+from typing import Generic, TypeVar
 
-from pydantic import BaseModel, ConfigDict
 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()
@@ -38,6 +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:
         """
@@ -66,7 +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
         """
@@ -91,7 +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.
@@ -142,7 +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:
@@ -151,13 +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,

server/reflector/processors/audio_chunker.py

@@ -1,27 +1,78 @@
-from reflector.processors.base import Processor
+from typing import Optional
+
 import av
+from prometheus_client import Counter, Histogram
+
+from reflector.processors.base import Processor
 
 
 class AudioChunkerProcessor(Processor):
     """
-    Assemble audio frames into chunks
+    Base class for assembling audio frames into chunks
     """
 
     INPUT_TYPE = av.AudioFrame
     OUTPUT_TYPE = list[av.AudioFrame]
 
-    def __init__(self, max_frames=256):
-        super().__init__()
+    m_chunk = Histogram(
+        "audio_chunker",
+        "Time spent in AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_call = Counter(
+        "audio_chunker_call",
+        "Number of calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_success = Counter(
+        "audio_chunker_success",
+        "Number of successful calls to AudioChunker.chunk",
+        ["backend"],
+    )
+    m_chunk_failure = Counter(
+        "audio_chunker_failure",
+        "Number of failed calls to AudioChunker.chunk",
+        ["backend"],
+    )
+
+    def __init__(self, *args, **kwargs):
+        name = self.__class__.__name__
+        self.m_chunk = self.m_chunk.labels(name)
+        self.m_chunk_call = self.m_chunk_call.labels(name)
+        self.m_chunk_success = self.m_chunk_success.labels(name)
+        self.m_chunk_failure = self.m_chunk_failure.labels(name)
+        super().__init__(*args, **kwargs)
         self.frames: list[av.AudioFrame] = []
-        self.max_frames = max_frames
 
     async def _push(self, data: av.AudioFrame):
-        self.frames.append(data)
-        if len(self.frames) >= self.max_frames:
-            await self.flush()
+        """Process incoming audio frame"""
+        # Validate audio format on first frame
+        if len(self.frames) == 0:
+            if data.sample_rate != 16000 or len(data.layout.channels) != 1:
+                raise ValueError(
+                    f"AudioChunkerProcessor expects 16kHz mono audio, got {data.sample_rate}Hz "
+                    f"with {len(data.layout.channels)} channel(s). "
+                    f"Use AudioDownscaleProcessor before this processor."
+                )
+
+        try:
+            self.m_chunk_call.inc()
+            with self.m_chunk.time():
+                result = await self._chunk(data)
+            self.m_chunk_success.inc()
+            if result:
+                await self.emit(result)
+        except Exception:
+            self.m_chunk_failure.inc()
+            raise
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """
+        Process audio frame and return chunk when ready.
+        Subclasses should implement their chunking logic here.
+        """
+        raise NotImplementedError
 
     async def _flush(self):
-        frames = self.frames[:]
-        self.frames = []
-        if frames:
-            await self.emit(frames)
+        """Flush any remaining frames when processing ends"""
+        raise NotImplementedError

server/reflector/processors/audio_chunker_auto.py

@@ -0,0 +1,32 @@
+import importlib
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.settings import settings
+
+
+class AudioChunkerAutoProcessor(AudioChunkerProcessor):
+    _registry = {}
+
+    @classmethod
+    def register(cls, name, kclass):
+        cls._registry[name] = kclass
+
+    def __new__(cls, name: str | None = None, **kwargs):
+        if name is None:
+            name = settings.AUDIO_CHUNKER_BACKEND
+        if name not in cls._registry:
+            module_name = f"reflector.processors.audio_chunker_{name}"
+            importlib.import_module(module_name)
+
+        # gather specific configuration for the processor
+        # search `AUDIO_CHUNKER_BACKEND_XXX_YYY`, push to constructor as `backend_xxx_yyy`
+        config = {}
+        name_upper = name.upper()
+        settings_prefix = "AUDIO_CHUNKER_"
+        config_prefix = f"{settings_prefix}{name_upper}_"
+        for key, value in settings:
+            if key.startswith(config_prefix):
+                config_name = key[len(settings_prefix) :].lower()
+                config[config_name] = value
+
+        return cls._registry[name](**config | kwargs)

server/reflector/processors/audio_chunker_frames.py

@@ -0,0 +1,34 @@
+from typing import Optional
+
+import av
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerFramesProcessor(AudioChunkerProcessor):
+    """
+    Simple frame-based audio chunker that emits chunks after a fixed number of frames
+    """
+
+    def __init__(self, max_frames=256, **kwargs):
+        super().__init__(**kwargs)
+        self.max_frames = max_frames
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        self.frames.append(data)
+        if len(self.frames) >= self.max_frames:
+            frames_to_emit = self.frames[:]
+            self.frames = []
+            return frames_to_emit
+
+        return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            await self.emit(frames)
+
+
+AudioChunkerAutoProcessor.register("frames", AudioChunkerFramesProcessor)

server/reflector/processors/audio_chunker_silero.py

@@ -0,0 +1,293 @@
+from typing import Optional
+
+import av
+import numpy as np
+import torch
+from silero_vad import VADIterator, load_silero_vad
+
+from reflector.processors.audio_chunker import AudioChunkerProcessor
+from reflector.processors.audio_chunker_auto import AudioChunkerAutoProcessor
+
+
+class AudioChunkerSileroProcessor(AudioChunkerProcessor):
+    """
+    Assemble audio frames into chunks with VAD-based speech detection using Silero VAD.
+
+    Expects input audio to be already downscaled to 16kHz mono s16 format
+    (handled by AudioDownscaleProcessor in the pipeline).
+    """
+
+    def __init__(
+        self,
+        block_frames=256,
+        max_frames=1024,
+        use_onnx=True,
+        min_frames=2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.block_frames = block_frames
+        self.max_frames = max_frames
+        self.min_frames = min_frames
+
+        # Initialize Silero VAD
+        self._init_vad(use_onnx)
+
+    def _init_vad(self, use_onnx=False):
+        """Initialize Silero VAD model for 16kHz audio"""
+        try:
+            torch.set_num_threads(1)
+            self.vad_model = load_silero_vad(onnx=use_onnx)
+            # VAD expects 16kHz audio (guaranteed by AudioDownscaleProcessor)
+            self.vad_iterator = VADIterator(self.vad_model, sampling_rate=16000)
+            self.logger.info("Silero VAD initialized for 16kHz audio")
+
+        except Exception as e:
+            self.logger.error(f"Failed to initialize Silero VAD: {e}")
+            self.vad_model = None
+            self.vad_iterator = None
+
+    async def _chunk(self, data: av.AudioFrame) -> Optional[list[av.AudioFrame]]:
+        """Process audio frame and return chunk when ready"""
+        self.frames.append(data)
+
+        # Check for speech segments every 32 frames (~1 second)
+        if len(self.frames) >= 32 and len(self.frames) % 32 == 0:
+            return await self._process_block()
+
+        # Safety fallback - emit if we hit max frames
+        elif len(self.frames) >= self.max_frames:
+            self.logger.warning(
+                f"AudioChunkerSileroProcessor: Reached max frames ({self.max_frames}), "
+                f"emitting first {self.max_frames // 2} frames"
+            )
+            frames_to_emit = self.frames[: self.max_frames // 2]
+            self.frames = self.frames[self.max_frames // 2 :]
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring fallback segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        return None
+
+    async def _process_block(self) -> Optional[list[av.AudioFrame]]:
+        # Need at least 32 frames for VAD detection (~1 second)
+        if len(self.frames) < 32 or self.vad_iterator is None:
+            return None
+
+        # Processing block with current buffer size
+        # print(f"Processing block: {len(self.frames)} frames in buffer")
+
+        try:
+            # Convert frames to numpy array for VAD
+            audio_array = self._frames_to_numpy(self.frames)
+
+            if audio_array is None:
+                # Fallback: emit all frames if conversion failed
+                frames_to_emit = self.frames[:]
+                self.frames = []
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring conversion-failed segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+                return None
+
+            # Find complete speech segments in the buffer
+            speech_end_frame = self._find_speech_segment_end(audio_array)
+
+            if speech_end_frame is None or speech_end_frame <= 0:
+                # No speech found but buffer is getting large
+                if len(self.frames) > 512:
+                    # Check if it's all silence and can be discarded
+                    # No speech segment found, buffer at {len(self.frames)} frames
+
+                    # Could emit silence or discard old frames here
+                    # For now, keep first 256 frames and discard older silence
+                    if len(self.frames) > 768:
+                        self.logger.debug(
+                            f"Discarding {len(self.frames) - 256} old frames (likely silence)"
+                        )
+                        self.frames = self.frames[-256:]
+                return None
+
+            # Calculate segment timing information
+            frames_to_emit = self.frames[:speech_end_frame]
+
+            # Get timing from av.AudioFrame
+            if frames_to_emit:
+                first_frame = frames_to_emit[0]
+                last_frame = frames_to_emit[-1]
+                sample_rate = first_frame.sample_rate
+
+                # Calculate duration
+                total_samples = sum(f.samples for f in frames_to_emit)
+                duration_seconds = total_samples / sample_rate if sample_rate > 0 else 0
+
+                # Get timestamps if available
+                start_time = (
+                    first_frame.pts * first_frame.time_base if first_frame.pts else 0
+                )
+                end_time = (
+                    last_frame.pts * last_frame.time_base if last_frame.pts else 0
+                )
+
+                # Convert to HH:MM:SS format for logging
+                def format_time(seconds):
+                    if not seconds:
+                        return "00:00:00"
+                    total_seconds = int(float(seconds))
+                    hours = total_seconds // 3600
+                    minutes = (total_seconds % 3600) // 60
+                    secs = total_seconds % 60
+                    return f"{hours:02d}:{minutes:02d}:{secs:02d}"
+
+                start_formatted = format_time(start_time)
+                end_formatted = format_time(end_time)
+
+                # Keep remaining frames for next processing
+                remaining_after = len(self.frames) - speech_end_frame
+
+                # Single structured log line
+                self.logger.info(
+                    "Speech segment found",
+                    start=start_formatted,
+                    end=end_formatted,
+                    frames=speech_end_frame,
+                    duration=round(duration_seconds, 2),
+                    buffer_before=len(self.frames),
+                    remaining=remaining_after,
+                )
+
+            # Keep remaining frames for next processing
+            self.frames = self.frames[speech_end_frame:]
+
+            # Filter out segments with too few frames
+            if len(frames_to_emit) >= self.min_frames:
+                return frames_to_emit
+            else:
+                self.logger.debug(
+                    f"Ignoring segment with {len(frames_to_emit)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+        except Exception as e:
+            self.logger.error(f"Error in VAD processing: {e}")
+            # Fallback to simple chunking
+            if len(self.frames) >= self.block_frames:
+                frames_to_emit = self.frames[: self.block_frames]
+                self.frames = self.frames[self.block_frames :]
+                if len(frames_to_emit) >= self.min_frames:
+                    return frames_to_emit
+                else:
+                    self.logger.debug(
+                        f"Ignoring exception-fallback segment with {len(frames_to_emit)} frames "
+                        f"(< {self.min_frames} minimum)"
+                    )
+
+        return None
+
+    def _frames_to_numpy(self, frames: list[av.AudioFrame]) -> Optional[np.ndarray]:
+        """Convert av.AudioFrame list to numpy array for VAD processing
+
+        Input frames are already 16kHz mono s16 format from AudioDownscaleProcessor.
+        Only need to convert s16 to float32 for Silero VAD.
+        """
+        if not frames:
+            return None
+
+        try:
+            # Concatenate all frame arrays
+            audio_arrays = [frame.to_ndarray().flatten() for frame in frames]
+            if not audio_arrays:
+                return None
+
+            combined_audio = np.concatenate(audio_arrays)
+
+            # Convert s16 to float32 (Silero VAD requires float32 in range [-1.0, 1.0])
+            # Input is guaranteed to be s16 from AudioDownscaleProcessor
+            return combined_audio.astype(np.float32) / 32768.0
+
+        except Exception as e:
+            self.logger.error(f"Error converting frames to numpy: {e}")
+            return None
+
+    def _find_speech_segment_end(self, audio_array: np.ndarray) -> Optional[int]:
+        """Find complete speech segments and return frame index at segment end"""
+        if self.vad_iterator is None or len(audio_array) == 0:
+            return None
+
+        try:
+            # Process audio in 512-sample windows for VAD
+            window_size = 512
+            min_silence_windows = 3  # Require 3 windows of silence after speech
+
+            # Track speech state
+            in_speech = False
+            speech_start = None
+            speech_end = None
+            silence_count = 0
+
+            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)))
+
+                # Detect if this window has speech
+                speech_dict = self.vad_iterator(chunk, return_seconds=True)
+
+                # VADIterator returns dict with 'start' and 'end' when speech segments are detected
+                if speech_dict:
+                    if not in_speech:
+                        # Speech started
+                        speech_start = i
+                        in_speech = True
+                        # Debug: print(f"Speech START at sample {i}, VAD: {speech_dict}")
+                    silence_count = 0  # Reset silence counter
+                    continue
+
+                if not in_speech:
+                    continue
+
+                # We're in speech but found silence
+                silence_count += 1
+                if silence_count < min_silence_windows:
+                    continue
+
+                # Found end of speech segment
+                speech_end = i - (min_silence_windows - 1) * window_size
+                # Debug: print(f"Speech END at sample {speech_end}")
+
+                # Convert sample position to frame index
+                samples_per_frame = self.frames[0].samples if self.frames else 1024
+                frame_index = speech_end // samples_per_frame
+
+                # Ensure we don't exceed buffer
+                frame_index = min(frame_index, len(self.frames))
+                return frame_index
+
+            return None
+
+        except Exception as e:
+            self.logger.error(f"Error finding speech segment: {e}")
+            return None
+
+    async def _flush(self):
+        frames = self.frames[:]
+        self.frames = []
+        if frames:
+            if len(frames) >= self.min_frames:
+                await self.emit(frames)
+            else:
+                self.logger.debug(
+                    f"Ignoring flush segment with {len(frames)} frames "
+                    f"(< {self.min_frames} minimum)"
+                )
+
+
+AudioChunkerAutoProcessor.register("silero", AudioChunkerSileroProcessor)

server/reflector/processors/audio_diarization.py

@@ -1,5 +1,10 @@
 from reflector.processors.base import Processor
-from reflector.processors.types import AudioDiarizationInput, TitleSummary, Word
+from reflector.processors.types import (
+    AudioDiarizationInput,
+    DiarizationSegment,
+    TitleSummary,
+    Word,
+)
 
 
 class AudioDiarizationProcessor(Processor):
@@ -33,18 +38,21 @@ class AudioDiarizationProcessor(Processor):
     async def _diarize(self, data: AudioDiarizationInput):
         raise NotImplementedError
 
-    def assign_speaker(self, words: list[Word], diarization: list[dict]):
-        self._diarization_remove_overlap(diarization)
-        self._diarization_remove_segment_without_words(words, diarization)
-        self._diarization_merge_same_speaker(words, diarization)
-        self._diarization_assign_speaker(words, diarization)
+    @classmethod
+    def assign_speaker(cls, words: list[Word], diarization: list[DiarizationSegment]):
+        cls._diarization_remove_overlap(diarization)
+        cls._diarization_remove_segment_without_words(words, diarization)
+        cls._diarization_merge_same_speaker(diarization)
+        cls._diarization_assign_speaker(words, diarization)
 
-    def iter_words_from_topics(self, topics: TitleSummary):
+    @staticmethod
+    def iter_words_from_topics(topics: list[TitleSummary]):
         for topic in topics:
             for word in topic.transcript.words:
                 yield word
 
-    def is_word_continuation(self, word_prev, word):
+    @staticmethod
+    def is_word_continuation(word_prev, word):
         """
         Return True if the word is a continuation of the previous word
         by checking if the previous word is ending with a punctuation
@@ -57,7 +65,8 @@ class AudioDiarizationProcessor(Processor):
             return False
         return True
 
-    def _diarization_remove_overlap(self, diarization: list[dict]):
+    @staticmethod
+    def _diarization_remove_overlap(diarization: list[DiarizationSegment]):
         """
         Remove overlap in diarization results
 
@@ -82,8 +91,9 @@ class AudioDiarizationProcessor(Processor):
             else:
                 diarization_idx += 1
 
+    @staticmethod
     def _diarization_remove_segment_without_words(
-        self, words: list[Word], diarization: list[dict]
+        words: list[Word], diarization: list[DiarizationSegment]
     ):
         """
         Remove diarization segments without words
@@ -112,9 +122,8 @@ class AudioDiarizationProcessor(Processor):
             else:
                 diarization_idx += 1
 
-    def _diarization_merge_same_speaker(
-        self, words: list[Word], diarization: list[dict]
-    ):
+    @staticmethod
+    def _diarization_merge_same_speaker(diarization: list[DiarizationSegment]):
         """
         Merge diarization contigous segments with the same speaker
 
@@ -131,7 +140,10 @@ class AudioDiarizationProcessor(Processor):
             else:
                 diarization_idx += 1
 
-    def _diarization_assign_speaker(self, words: list[Word], diarization: list[dict]):
+    @classmethod
+    def _diarization_assign_speaker(
+        cls, words: list[Word], diarization: list[DiarizationSegment]
+    ):
         """
         Assign speaker to words based on diarization
 
@@ -139,7 +151,7 @@ class AudioDiarizationProcessor(Processor):
         """
 
         word_idx = 0
-        last_speaker = None
+        last_speaker = 0
         for d in diarization:
             start = d["start"]
             end = d["end"]
@@ -154,7 +166,7 @@ class AudioDiarizationProcessor(Processor):
                     # If it's a continuation, assign with the last speaker
                     is_continuation = False
                     if word_idx > 0 and word_idx < len(words) - 1:
-                        is_continuation = self.is_word_continuation(
+                        is_continuation = cls.is_word_continuation(
                             *words[word_idx - 1 : word_idx + 1]
                         )
                     if is_continuation:

server/reflector/processors/audio_diarization_modal.py

@@ -1,4 +1,5 @@
 import httpx
+
 from reflector.processors.audio_diarization import AudioDiarizationProcessor
 from reflector.processors.audio_diarization_auto import AudioDiarizationAutoProcessor
 from reflector.processors.types import AudioDiarizationInput, TitleSummary
@@ -9,12 +10,17 @@ class AudioDiarizationModalProcessor(AudioDiarizationProcessor):
     INPUT_TYPE = AudioDiarizationInput
     OUTPUT_TYPE = TitleSummary
 
-    def __init__(self, **kwargs):
+    def __init__(self, modal_api_key: str | None = None, **kwargs):
         super().__init__(**kwargs)
+        if not settings.DIARIZATION_URL:
+            raise Exception(
+                "DIARIZATION_URL required to use AudioDiarizationModalProcessor"
+            )
         self.diarization_url = settings.DIARIZATION_URL + "/diarize"
-        self.headers = {
-            "Authorization": f"Bearer {settings.LLM_MODAL_API_KEY}",
-        }
+        self.modal_api_key = modal_api_key
+        self.headers = {}
+        if self.modal_api_key:
+            self.headers["Authorization"] = f"Bearer {self.modal_api_key}"
 
     async def _diarize(self, data: AudioDiarizationInput):
         # Gather diarization data

server/reflector/processors/audio_diarization_pyannote.py

@@ -0,0 +1,74 @@
+import os
+
+import torch
+import torchaudio
+from pyannote.audio import Pipeline
+
+from reflector.processors.audio_diarization import AudioDiarizationProcessor
+from reflector.processors.audio_diarization_auto import AudioDiarizationAutoProcessor
+from reflector.processors.types import AudioDiarizationInput, DiarizationSegment
+
+
+class AudioDiarizationPyannoteProcessor(AudioDiarizationProcessor):
+    """Local diarization processor using pyannote.audio library"""
+
+    def __init__(
+        self,
+        model_name: str = "pyannote/speaker-diarization-3.1",
+        pyannote_auth_token: str | None = None,
+        device: str | None = None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.model_name = model_name
+        self.auth_token = pyannote_auth_token or os.environ.get("HF_TOKEN")
+        self.device = device
+
+        if device is None:
+            self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        self.logger.info(f"Loading pyannote diarization model: {self.model_name}")
+        self.diarization_pipeline = Pipeline.from_pretrained(
+            self.model_name, use_auth_token=self.auth_token
+        )
+        self.diarization_pipeline.to(torch.device(self.device))
+        self.logger.info(f"Diarization model loaded on device: {self.device}")
+
+    async def _diarize(self, data: AudioDiarizationInput) -> list[DiarizationSegment]:
+        try:
+            # Load audio file (audio_url is assumed to be a local file path)
+            self.logger.info(f"Loading local audio file: {data.audio_url}")
+            waveform, sample_rate = torchaudio.load(data.audio_url)
+            audio_input = {"waveform": waveform, "sample_rate": sample_rate}
+            self.logger.info("Running speaker diarization")
+            diarization = self.diarization_pipeline(audio_input)
+
+            # Convert pyannote diarization output to our format
+            segments = []
+            for segment, _, speaker in diarization.itertracks(yield_label=True):
+                # Extract speaker number from label (e.g., "SPEAKER_00" -> 0)
+                speaker_id = 0
+                if speaker.startswith("SPEAKER_"):
+                    try:
+                        speaker_id = int(speaker.split("_")[-1])
+                    except (ValueError, IndexError):
+                        # Fallback to hash-based ID if parsing fails
+                        speaker_id = hash(speaker) % 1000
+
+                segments.append(
+                    {
+                        "start": round(segment.start, 3),
+                        "end": round(segment.end, 3),
+                        "speaker": speaker_id,
+                    }
+                )
+
+            self.logger.info(f"Diarization completed with {len(segments)} segments")
+            return segments
+
+        except Exception as e:
+            self.logger.exception(f"Diarization failed: {e}")
+            raise
+
+
+AudioDiarizationAutoProcessor.register("pyannote", AudioDiarizationPyannoteProcessor)

server/reflector/processors/audio_downscale.py

@@ -0,0 +1,60 @@
+from typing import Optional
+
+import av
+from av.audio.resampler import AudioResampler
+
+from reflector.processors.base import Processor
+
+
+def copy_frame(frame: av.AudioFrame) -> av.AudioFrame:
+    frame_copy = frame.from_ndarray(
+        frame.to_ndarray(),
+        format=frame.format.name,
+        layout=frame.layout.name,
+    )
+    frame_copy.sample_rate = frame.sample_rate
+    frame_copy.pts = frame.pts
+    frame_copy.time_base = frame.time_base
+    return frame_copy
+
+
+class AudioDownscaleProcessor(Processor):
+    """
+    Downscale audio frames to 16kHz mono format
+    """
+
+    INPUT_TYPE = av.AudioFrame
+    OUTPUT_TYPE = av.AudioFrame
+
+    def __init__(self, target_rate: int = 16000, target_layout: str = "mono", **kwargs):
+        super().__init__(**kwargs)
+        self.target_rate = target_rate
+        self.target_layout = target_layout
+        self.resampler: Optional[AudioResampler] = None
+        self.needs_resampling: Optional[bool] = None
+
+    async def _push(self, data: av.AudioFrame):
+        if self.needs_resampling is None:
+            self.needs_resampling = (
+                data.sample_rate != self.target_rate
+                or data.layout.name != self.target_layout
+            )
+
+            if self.needs_resampling:
+                self.resampler = AudioResampler(
+                    format="s16", layout=self.target_layout, rate=self.target_rate
+                )
+
+        if not self.needs_resampling or not self.resampler:
+            await self.emit(data)
+            return
+
+        resampled_frames = self.resampler.resample(copy_frame(data))
+        for resampled_frame in resampled_frames:
+            await self.emit(resampled_frame)
+
+    async def _flush(self):
+        if self.needs_resampling and self.resampler:
+            final_frames = self.resampler.resample(None)
+            for frame in final_frames:
+                await self.emit(frame)

server/reflector/processors/audio_file_writer.py

@@ -1,6 +1,7 @@
 from pathlib import Path
 
 import av
+
 from reflector.processors.base import Processor
 
 

server/reflector/processors/audio_merge.py

@@ -1,10 +1,12 @@
-from reflector.processors.base import Processor
-from reflector.processors.types import AudioFile
+import io
 from time import monotonic_ns
 from uuid import uuid4
-import io
+
 import av
 
+from reflector.processors.base import Processor
+from reflector.processors.types import AudioFile
+
 
 class AudioMergeProcessor(Processor):
     """
@@ -14,37 +16,46 @@ class AudioMergeProcessor(Processor):
     INPUT_TYPE = list[av.AudioFrame]
     OUTPUT_TYPE = AudioFile
 
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
     async def _push(self, data: list[av.AudioFrame]):
         if not data:
             return
 
         # get audio information from first frame
         frame = data[0]
-        channels = len(frame.layout.channels)
-        sample_rate = frame.sample_rate
-        sample_width = frame.format.bytes
+        output_channels = len(frame.layout.channels)
+        output_sample_rate = frame.sample_rate
+        output_sample_width = frame.format.bytes
 
         # create audio file
         uu = uuid4().hex
         fd = io.BytesIO()
 
+        # Use PyAV to write frames
         out_container = av.open(fd, "w", format="wav")
-        out_stream = out_container.add_stream("pcm_s16le", rate=sample_rate)
+        out_stream = out_container.add_stream("pcm_s16le", rate=output_sample_rate)
+        out_stream.layout = frame.layout.name
+
         for frame in data:
             for packet in out_stream.encode(frame):
                 out_container.mux(packet)
+
+        # Flush the encoder
         for packet in out_stream.encode(None):
             out_container.mux(packet)
         out_container.close()
+
         fd.seek(0)
 
         # emit audio file
         audiofile = AudioFile(
             name=f"{monotonic_ns()}-{uu}.wav",
             fd=fd,
-            sample_rate=sample_rate,
-            channels=channels,
-            sample_width=sample_width,
+            sample_rate=output_sample_rate,
+            channels=output_channels,
+            sample_width=output_sample_width,
             timestamp=data[0].pts * data[0].time_base,
         )
 

server/reflector/processors/audio_transcript.py

@@ -1,4 +1,5 @@
 from prometheus_client import Counter, Histogram
+
 from reflector.processors.base import Processor
 from reflector.processors.types import AudioFile, Transcript
 

server/reflector/processors/audio_transcript_modal.py

@@ -13,6 +13,7 @@ API will be a POST request to TRANSCRIPT_URL:
 """
 
 from openai import AsyncOpenAI
+
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
 from reflector.processors.audio_transcript_auto import AudioTranscriptAutoProcessor
 from reflector.processors.types import AudioFile, Transcript, Word
@@ -20,16 +21,24 @@ from reflector.settings import settings
 
 
 class AudioTranscriptModalProcessor(AudioTranscriptProcessor):
-    def __init__(self, modal_api_key: str):
+    def __init__(
+        self,
+        modal_api_key: str | None = None,
+        **kwargs,
+    ):
         super().__init__()
+        if not settings.TRANSCRIPT_URL:
+            raise Exception(
+                "TRANSCRIPT_URL required to use AudioTranscriptModalProcessor"
+            )
         self.transcript_url = settings.TRANSCRIPT_URL + "/v1"
         self.timeout = settings.TRANSCRIPT_TIMEOUT
-        self.api_key = settings.TRANSCRIPT_MODAL_API_KEY
+        self.modal_api_key = modal_api_key
 
     async def _transcript(self, data: AudioFile):
         async with AsyncOpenAI(
             base_url=self.transcript_url,
-            api_key=self.api_key,
+            api_key=self.modal_api_key,
             timeout=self.timeout,
         ) as client:
             self.logger.debug(f"Try to transcribe audio {data.name}")

server/reflector/processors/audio_transcript_whisper.py

@@ -1,4 +1,5 @@
 from faster_whisper import WhisperModel
+
 from reflector.processors.audio_transcript import AudioTranscriptProcessor
 from reflector.processors.audio_transcript_auto import AudioTranscriptAutoProcessor
 from reflector.processors.types import AudioFile, Transcript, Word

server/reflector/processors/base.py

@@ -5,6 +5,7 @@ from uuid import uuid4
 
 from prometheus_client import Counter, Gauge, Histogram
 from pydantic import BaseModel
+
 from reflector.logger import logger
 
 
@@ -172,6 +173,7 @@ class Processor(Emitter):
         except Exception:
             self.m_processor_failure.inc()
             self.logger.exception("Error in push")
+            raise
 
     async def flush(self):
         """
@@ -239,33 +241,45 @@ class ThreadedProcessor(Processor):
         self.INPUT_TYPE = processor.INPUT_TYPE
         self.OUTPUT_TYPE = processor.OUTPUT_TYPE
         self.executor = ThreadPoolExecutor(max_workers=max_workers)
-        self.queue = asyncio.Queue()
-        self.task = asyncio.get_running_loop().create_task(self.loop())
+        self.queue = asyncio.Queue(maxsize=50)
+        self.task: asyncio.Task | None = None
 
     def set_pipeline(self, pipeline: "Pipeline"):
         super().set_pipeline(pipeline)
         self.processor.set_pipeline(pipeline)
 
     async def loop(self):
-        while True:
-            data = await self.queue.get()
-            self.m_processor_queue.set(self.queue.qsize())
-            with self.m_processor_queue_in_progress.track_inprogress():
-                try:
-                    if data is None:
-                        await self.processor.flush()
-                        break
+        try:
+            while True:
+                data = await self.queue.get()
+                self.m_processor_queue.set(self.queue.qsize())
+                with self.m_processor_queue_in_progress.track_inprogress():
                     try:
-                        await self.processor.push(data)
-                    except Exception:
-                        self.logger.error(
-                            f"Error in push {self.processor.__class__.__name__}"
-                            ", continue"
-                        )
-                finally:
-                    self.queue.task_done()
+                        if data is None:
+                            await self.processor.flush()
+                            break
+                        try:
+                            await self.processor.push(data)
+                        except Exception:
+                            self.logger.error(
+                                f"Error in push {self.processor.__class__.__name__}"
+                                ", continue"
+                            )
+                    finally:
+                        self.queue.task_done()
+        except Exception as e:
+            logger.error(f"Crash in {self.__class__.__name__}: {e}", exc_info=e)
+
+    async def _ensure_task(self):
+        if self.task is None:
+            self.task = asyncio.get_running_loop().create_task(self.loop())
+
+        # XXX not doing a sleep here make the whole pipeline prior the thread
+        # to be running without having a chance to work on the task here.
+        await asyncio.sleep(0)
 
     async def _push(self, data):
+        await self._ensure_task()
         await self.queue.put(data)
 
     async def _flush(self):

server/reflector/processors/file_diarization.py

@@ -0,0 +1,33 @@
+from pydantic import BaseModel
+
+from reflector.processors.base import Processor
+from reflector.processors.types import DiarizationSegment
+
+
+class FileDiarizationInput(BaseModel):
+    """Input for file diarization containing audio URL"""
+
+    audio_url: str
+
+
+class FileDiarizationOutput(BaseModel):
+    """Output for file diarization containing speaker segments"""
+
+    diarization: list[DiarizationSegment]
+
+
+class FileDiarizationProcessor(Processor):
+    """
+    Diarize complete audio files from URL
+    """
+
+    INPUT_TYPE = FileDiarizationInput
+    OUTPUT_TYPE = FileDiarizationOutput
+
+    async def _push(self, data: FileDiarizationInput):
+        result = await self._diarize(data)
+        if result:
+            await self.emit(result)
+
+    async def _diarize(self, data: FileDiarizationInput):
+        raise NotImplementedError

server/reflector/processors/file_diarization_auto.py

@@ -0,0 +1,33 @@
+import importlib
+
+from reflector.processors.file_diarization import FileDiarizationProcessor
+from reflector.settings import settings
+
+
+class FileDiarizationAutoProcessor(FileDiarizationProcessor):
+    _registry = {}
+
+    @classmethod
+    def register(cls, name, kclass):
+        cls._registry[name] = kclass
+
+    def __new__(cls, name: str | None = None, **kwargs):
+        if name is None:
+            name = settings.DIARIZATION_BACKEND
+
+        if name not in cls._registry:
+            module_name = f"reflector.processors.file_diarization_{name}"
+            importlib.import_module(module_name)
+
+        # gather specific configuration for the processor
+        # search `DIARIZATION_BACKEND_XXX_YYY`, push to constructor as `backend_xxx_yyy`
+        config = {}
+        name_upper = name.upper()
+        settings_prefix = "DIARIZATION_"
+        config_prefix = f"{settings_prefix}{name_upper}_"
+        for key, value in settings:
+            if key.startswith(config_prefix):
+                config_name = key[len(settings_prefix) :].lower()
+                config[config_name] = value
+
+        return cls._registry[name](**config | kwargs)

server/reflector/processors/file_diarization_modal.py

@@ -0,0 +1,57 @@
+"""
+File diarization implementation using the GPU service from modal.com
+
+API will be a POST request to DIARIZATION_URL:
+
+```
+POST /diarize?audio_file_url=...&timestamp=0
+Authorization: Bearer <modal_api_key>
+```
+"""
+
+import httpx
+
+from reflector.processors.file_diarization import (
+    FileDiarizationInput,
+    FileDiarizationOutput,
+    FileDiarizationProcessor,
+)
+from reflector.processors.file_diarization_auto import FileDiarizationAutoProcessor
+from reflector.settings import settings
+
+
+class FileDiarizationModalProcessor(FileDiarizationProcessor):
+    def __init__(self, modal_api_key: str | None = None, **kwargs):
+        super().__init__(**kwargs)
+        if not settings.DIARIZATION_URL:
+            raise Exception(
+                "DIARIZATION_URL required to use FileDiarizationModalProcessor"
+            )
+        self.diarization_url = settings.DIARIZATION_URL + "/diarize"
+        self.file_timeout = settings.DIARIZATION_FILE_TIMEOUT
+        self.modal_api_key = modal_api_key
+
+    async def _diarize(self, data: FileDiarizationInput):
+        """Get speaker diarization for file"""
+        self.logger.info(f"Starting diarization from {data.audio_url}")
+
+        headers = {}
+        if self.modal_api_key:
+            headers["Authorization"] = f"Bearer {self.modal_api_key}"
+
+        async with httpx.AsyncClient(timeout=self.file_timeout) as client:
+            response = await client.post(
+                self.diarization_url,
+                headers=headers,
+                params={
+                    "audio_file_url": data.audio_url,
+                    "timestamp": 0,
+                },
+            )
+            response.raise_for_status()
+            diarization_data = response.json()["diarization"]
+
+        return FileDiarizationOutput(diarization=diarization_data)
+
+
+FileDiarizationAutoProcessor.register("modal", FileDiarizationModalProcessor)

server/reflector/processors/file_transcript.py

@@ -0,0 +1,65 @@
+from prometheus_client import Counter, Histogram
+
+from reflector.processors.base import Processor
+from reflector.processors.types import Transcript
+
+
+class FileTranscriptInput:
+    """Input for file transcription containing audio URL and language settings"""
+
+    def __init__(self, audio_url: str, language: str = "en"):
+        self.audio_url = audio_url
+        self.language = language
+
+
+class FileTranscriptProcessor(Processor):
+    """
+    Transcript complete audio files from URL
+    """
+
+    INPUT_TYPE = FileTranscriptInput
+    OUTPUT_TYPE = Transcript
+
+    m_transcript = Histogram(
+        "file_transcript",
+        "Time spent in FileTranscript.transcript",
+        ["backend"],
+    )
+    m_transcript_call = Counter(
+        "file_transcript_call",
+        "Number of calls to FileTranscript.transcript",
+        ["backend"],
+    )
+    m_transcript_success = Counter(
+        "file_transcript_success",
+        "Number of successful calls to FileTranscript.transcript",
+        ["backend"],
+    )
+    m_transcript_failure = Counter(
+        "file_transcript_failure",
+        "Number of failed calls to FileTranscript.transcript",
+        ["backend"],
+    )
+
+    def __init__(self, *args, **kwargs):
+        name = self.__class__.__name__
+        self.m_transcript = self.m_transcript.labels(name)
+        self.m_transcript_call = self.m_transcript_call.labels(name)
+        self.m_transcript_success = self.m_transcript_success.labels(name)
+        self.m_transcript_failure = self.m_transcript_failure.labels(name)
+        super().__init__(*args, **kwargs)
+
+    async def _push(self, data: FileTranscriptInput):
+        try:
+            self.m_transcript_call.inc()
+            with self.m_transcript.time():
+                result = await self._transcript(data)
+            self.m_transcript_success.inc()
+            if result:
+                await self.emit(result)
+        except Exception:
+            self.m_transcript_failure.inc()
+            raise
+
+    async def _transcript(self, data: FileTranscriptInput):
+        raise NotImplementedError