selfhostyourtech/reflector
2
0
Fork 0
mirror of https://github.com/Monadical-SAS/reflector.git synced 2025-12-20 20:29:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

move pipeline dev docs to dev docs location

Browse Source
This commit is contained in:
Igor Loskutov
2025-12-12 16:20:11 -05:00
parent fd5298c1ee
commit ba8568752e
2 changed files with 422 additions and 418 deletions
Download Patch File Download Diff File Expand all files Collapse all files

419
docs/docs/pipelines/live-pipeline.md
View File

@@ -4,421 +4,4 @@ title: Live pipeline
# Live pipeline
# Daily.co
This document details every external call, storage operation, and database write that occurs when a new Daily.co recording is discovered.
## Trigger
Two entry points, both converging to the same handler:
1. **Webhook**: Daily.co sends `POST /v1/daily/webhook` with `recording.ready-to-download`
2. **Polling**: `GET /recordings` (paginated, max 100/call) → filter new → convert to same payload format
Both produce `RecordingReadyPayload` and call `handleRecordingReady(payload)`.
```
┌─────────────────┐ ┌──────────────────────────┐
│ Daily Webhook │────▶│ RecordingReadyPayload │
│ (push) │ │ {room_name, recording_id│
└─────────────────┘ │ tracks[], ...} │
└────────────┬─────────────┘
┌─────────────────┐ │
│ GET /recordings│ ▼
│ (poll) │────▶ convert ──▶ handleRecordingReady()
└─────────────────┘ │
▼
┌────────────────────────┐
│ process_multitrack_ │
│ recording pipeline │
└────────────────────────┘
```
**Polling API**: `GET https://api.daily.co/v1/recordings`
- Pagination: `limit` (max 100), `starting_after`, `ending_before`
- Rate limit: ~2 req/sec
- Response: `{total_count, data: Recording[]}`
```mermaid
flowchart TB
subgraph Trigger["1. Recording Discovery - Daily.co Webhook"]
DAILY_WEBHOOK["Daily.co sends POST /v1/daily/webhook<br/>type: recording.ready-to-download"]
VERIFY["Verify X-Webhook-Signature (HMAC)"]
PARSE["Parse DailyWebhookEvent<br/>Extract tracks[], room_name, recording_id"]
FILTER["Filter audio tracks only<br/>track_keys = [t.s3Key for t in tracks if t.type == 'audio']"]
DISPATCH["process_multitrack_recording.delay()"]
DAILY_WEBHOOK --> VERIFY --> PARSE --> FILTER --> DISPATCH
end
subgraph Init["2. Recording Initialization"]
FETCH_MEETING[DB READ: meetings_controller.get_by_room_name]
FETCH_ROOM[DB READ: rooms_controller.get_by_name]
DAILY_API_REC[Daily API: GET /recordings/recording_id]
DAILY_API_PART[Daily API: GET /meetings/mtgSessionId/participants]
CREATE_RECORDING[DB WRITE: recordings_controller.create]
CREATE_TRANSCRIPT[DB WRITE: transcripts_controller.add]
MAP_PARTICIPANTS[DB WRITE: transcript.participants upsert]
end
subgraph Pipeline["3. Processing Pipeline"]
direction TB
PAD[Track Padding & Mixdown]
TRANSCRIBE[GPU: Transcription per track]
TOPICS[LLM: Topic Detection]
TITLE[LLM: Title Generation]
SUMMARY[LLM: Summary Generation]
end
subgraph Storage["4. S3 Operations"]
S3_PRESIGN[S3: generate_presigned_url for tracks]
S3_UPLOAD_PADDED[S3 UPLOAD: padded tracks temp]
S3_UPLOAD_MP3[S3 UPLOAD: audio.mp3]
S3_DELETE_TEMP[S3 DELETE: cleanup temp files]
end
subgraph PostProcess["5. Post-Processing"]
CONSENT[Consent check & cleanup]
ZULIP[Zulip: send/update message]
WEBHOOK_OUT[Webhook: POST to room.webhook_url]
end
Trigger --> Init --> Pipeline
Pipeline --> Storage
Pipeline --> PostProcess
```
## Detailed Sequence: Daily.co Multitrack Recording
```mermaid
sequenceDiagram
participant DailyCo as Daily.co
participant API as FastAPI /v1/daily/webhook
participant Worker as Celery Worker
participant DB as PostgreSQL
participant DailyAPI as Daily.co REST API
participant S3 as AWS S3
participant GPU as Modal.com GPU
participant LLM as LLM Service
participant WS as WebSocket
participant Zulip as Zulip
participant ExtWH as External Webhook
Note over DailyCo,API: Phase 0: Webhook Receipt
DailyCo->>API: POST /v1/daily/webhook
Note right of DailyCo: X-Webhook-Signature, X-Webhook-Timestamp
API->>API: verify_webhook_signature()
API->>API: Extract audio track s3Keys from payload.tracks[]
API->>Worker: process_multitrack_recording.delay()
API-->>DailyCo: 200 OK
Note over Worker,DailyAPI: Phase 1: Recording Initialization
Worker->>DB: SELECT meeting WHERE room_name=?
Worker->>DB: SELECT room WHERE name=?
Worker->>DailyAPI: GET /recordings/{recording_id}
DailyAPI-->>Worker: {mtgSessionId, ...}
Worker->>DailyAPI: GET /meetings/{mtgSessionId}/participants
DailyAPI-->>Worker: [{participant_id, user_name}, ...]
Worker->>DB: INSERT INTO recording
Worker->>DB: INSERT INTO transcript (status='idle')
loop For each track_key (parse participant_id from filename)
Worker->>DB: UPSERT transcript.participants[speaker=idx, name=X]
end
Note over Worker,S3: Phase 2: Track Padding
Worker->>DB: UPDATE transcript SET status='processing'
Worker->>WS: broadcast STATUS='processing'
loop For each track in track_keys (N tracks)
Worker->>S3: generate_presigned_url(track_key, DAILYCO_BUCKET)
S3-->>Worker: presigned_url (2hr)
Note over Worker: PyAV: read WebM, extract start_time
Note over Worker: PyAV: adelay filter (pad silence)
Worker->>S3: PUT file_pipeline/{id}/tracks/padded_{idx}.webm
Worker->>S3: generate_presigned_url(padded_{idx}.webm)
end
Note over Worker,S3: Phase 3: Audio Mixdown
Note over Worker: PyAV: amix filter → stereo MP3
Worker->>DB: UPDATE transcript SET duration=X
Worker->>WS: broadcast DURATION
Worker->>S3: PUT {transcript_id}/audio.mp3
Worker->>DB: UPDATE transcript SET audio_location='storage'
Note over Worker: Phase 4: Waveform
Note over Worker: Generate peaks from MP3
Worker->>DB: UPDATE events+=WAVEFORM
Worker->>WS: broadcast WAVEFORM
Note over Worker,GPU: Phase 5: Transcription (N GPU calls)
loop For each padded track URL (N tracks)
Worker->>GPU: POST /v1/audio/transcriptions-from-url
Note right of GPU: {audio_file_url, language, batch:true}
GPU-->>Worker: {words: [{word, start, end}, ...]}
Note over Worker: Assign speaker=track_idx to words
end
Note over Worker: Merge all words, sort by start time
Worker->>DB: UPDATE events+=TRANSCRIPT
Worker->>WS: broadcast TRANSCRIPT
Note over Worker,S3: Cleanup temp files
loop For each padded file
Worker->>S3: DELETE padded_{idx}.webm
end
Note over Worker,LLM: Phase 6: Topic Detection (C LLM calls)
Note over Worker: C = ceil(total_words / 300)
loop For each 300-word chunk (C chunks)
Worker->>LLM: TOPIC_PROMPT + words[i:i+300]
Note right of LLM: "Extract main topic title + 2-sentence summary"
LLM-->>Worker: TitleSummary{title, summary}
Worker->>DB: UPSERT topics[]
Worker->>DB: UPDATE events+=TOPIC
Worker->>WS: broadcast TOPIC
end
Note over Worker,LLM: Phase 7a: Title Generation (1 LLM call)
Note over Worker: Input: all TitleSummary[].title joined
Worker->>LLM: TITLE_PROMPT
Note right of LLM: "Generate concise title from topic titles"
LLM-->>Worker: "Meeting Title"
Worker->>DB: UPDATE transcript SET title=X
Worker->>DB: UPDATE events+=FINAL_TITLE
Worker->>WS: broadcast FINAL_TITLE
Note over Worker,LLM: Phase 7b: Summary Generation (2+2M LLM calls)
Note over Worker: Reconstruct full transcript from TitleSummary[].transcript
opt If participants unknown
Worker->>LLM: PARTICIPANTS_PROMPT
LLM-->>Worker: ParticipantsResponse
end
Worker->>LLM: SUBJECTS_PROMPT (call #1)
Note right of LLM: "Main high-level topics? Max 6"
LLM-->>Worker: SubjectsResponse{subjects: ["A", "B", ...]}
loop For each subject (M subjects, max 6)
Worker->>LLM: DETAILED_SUBJECT_PROMPT (call #2..#1+M)
Note right of LLM: "Info about 'A': decisions, actions, deadlines"
LLM-->>Worker: detailed_response (discarded after next call)
Worker->>LLM: PARAGRAPH_SUMMARY_PROMPT (call #2+M..#1+2M)
Note right of LLM: "Summarize in 1 paragraph"
LLM-->>Worker: paragraph → summaries[]
end
Worker->>LLM: RECAP_PROMPT (call #2+2M)
Note right of LLM: "High-level quick recap, 1 paragraph"
LLM-->>Worker: recap
Note over Worker: long_summary = "# Quick recap\n{recap}\n# Summary\n**A**\n{para1}..."
Note over Worker: short_summary = recap only
Worker->>DB: UPDATE long_summary, short_summary
Worker->>DB: UPDATE events+=FINAL_LONG_SUMMARY
Worker->>WS: broadcast FINAL_LONG_SUMMARY
Worker->>DB: UPDATE events+=FINAL_SHORT_SUMMARY
Worker->>WS: broadcast FINAL_SHORT_SUMMARY
Note over Worker,DB: Phase 8: Finalize
Worker->>DB: UPDATE transcript SET status='ended'
Worker->>DB: UPDATE events+=STATUS
Worker->>WS: broadcast STATUS='ended'
Note over Worker,ExtWH: Phase 9: Post-Processing Chain
Worker->>DB: SELECT meeting_consent WHERE meeting_id=?
alt Any consent denied
Worker->>S3: DELETE tracks from DAILYCO_BUCKET
Worker->>S3: DELETE audio.mp3 from TRANSCRIPT_BUCKET
Worker->>DB: UPDATE transcript SET audio_deleted=true
end
opt Room has zulip_auto_post=true
alt Existing zulip_message_id
Worker->>Zulip: PATCH /api/v1/messages/{id}
else New
Worker->>Zulip: POST /api/v1/messages
Zulip-->>Worker: {id}
Worker->>DB: UPDATE transcript SET zulip_message_id=X
end
end
opt Room has webhook_url
Worker->>ExtWH: POST {webhook_url}
Note right of ExtWH: X-Webhook-Signature: HMAC-SHA256
Note right of ExtWH: Body: {transcript_id, room_id, ...}
end
```
## Title & Summary Generation Data Flow
```mermaid
flowchart TB
subgraph Input["Input: TitleSummary[] from Topic Detection"]
TS1["TitleSummary 1<br/>title: 'Q1 Budget'<br/>transcript: words[0:300]"]
TS2["TitleSummary 2<br/>title: 'Product Launch'<br/>transcript: words[300:600]"]
TS3["TitleSummary N..."]
end
subgraph TitleGen["Title Generation"]
T1["Extract .title from each TitleSummary"]
T2["Concatenate: '- Q1 Budget\n- Product Launch\n...'"]
T3["LLM: TITLE_PROMPT\n'Generate concise title from topic titles'"]
T4["Output: FinalTitle"]
T1 --> T2 --> T3 --> T4
end
subgraph SummaryGen["Summary Generation"]
direction TB
subgraph Reconstruct["1. Reconstruct Full Transcript"]
S1["For each TitleSummary.transcript.as_segments()"]
S2["Map speaker ID → name"]
S3["Build: 'Alice: hello\nBob: hi\n...'"]
S1 --> S2 --> S3
end
subgraph Subjects["2. Extract Subjects - LLM call #1"]
S4["LLM: SUBJECTS_PROMPT\n'Main high-level topics? Max 6'"]
S5["subjects[] = ['Budget Review', ...]"]
S4 --> S5
end
subgraph DetailedSum["3. Per-Subject Summary - LLM calls #2 to #(1+2M)"]
S6["For each subject:"]
S7["LLM: DETAILED_SUBJECT_PROMPT\n'Info about subject: decisions, actions...'"]
S8["detailed_response - NOT STORED"]
S9["LLM: PARAGRAPH_SUMMARY_PROMPT\n'Summarize in 1 paragraph'"]
S10["paragraph → summaries[]"]
S6 --> S7 --> S8 --> S9 --> S10
end
subgraph Recap["4. Generate Recap - LLM call #(2+2M)"]
S11["Concatenate paragraph summaries"]
S12["LLM: RECAP_PROMPT\n'High-level recap, 1 paragraph'"]
S13["recap"]
S11 --> S12 --> S13
end
subgraph Output["5. Output"]
S14["long_summary = markdown:\n# Quick recap\n[recap]\n# Summary\n**Subject 1**\n[para1]..."]
S15["short_summary = recap only"]
S14 --> S15
end
Reconstruct --> Subjects --> DetailedSum --> Recap --> Output
end
Input --> TitleGen
Input --> SummaryGen
```
### topics[] vs subjects[]
| | topics[] | subjects[] |
|-|----------|------------|
| **Source** | 300-word chunk splitting | LLM extraction from full text |
| **Count** | Variable (words / 300) | Max 6 |
| **Purpose** | Timeline segmentation | Summary structure |
| **Has timestamp?** | Yes | No |
## External API Calls Summary
### 1. Daily.co REST API (called during initialization)
| Endpoint | Method | When | Purpose |
|----------|--------|------|---------|
| `GET /recordings/{recording_id}` | GET | After webhook | Get mtgSessionId for participant lookup |
| `GET /meetings/{mtgSessionId}/participants` | GET | After above | Map participant_id → user_name |
### 2. GPU Service (Modal.com or Self-Hosted)
| Endpoint | Method | Count | Request |
|----------|--------|-------|---------|
| `{TRANSCRIPT_URL}/v1/audio/transcriptions-from-url` | POST | **N** (N = num tracks) | `{audio_file_url, language, batch: true}` |
**Note**: Diarization is NOT called for multitrack - speaker identification comes from separate tracks.
### 3. LLM Service (OpenAI-compatible via LlamaIndex)
| Phase | Operation | Input | LLM Calls | Output |
|-------|-----------|-------|-----------|--------|
| Topic Detection | TOPIC_PROMPT per 300-word chunk | words[i:i+300] | **C** = ceil(words/300) | TitleSummary{title, summary, timestamp} |
| Title Generation | TITLE_PROMPT | All topic titles joined | **1** | FinalTitle |
| Participant ID | PARTICIPANTS_PROMPT | Full transcript | **0-1** (skipped if known) | ParticipantsResponse |
| Subject Extraction | SUBJECTS_PROMPT | Full transcript | **1** | SubjectsResponse{subjects[]} |
| Subject Detail | DETAILED_SUBJECT_PROMPT | Full transcript + subject name | **M** (M = subjects, max 6) | detailed text (discarded) |
| Subject Paragraph | PARAGRAPH_SUMMARY_PROMPT | Detailed text | **M** | paragraph text → summaries[] |
| Recap | RECAP_PROMPT | All paragraph summaries | **1** | recap text |
**Total LLM calls**: C + 2M + 3 (+ 1 if participants unknown)
- Short meeting (1000 words, 3 subjects): ~4 + 6 + 3 = **13 calls**
- Long meeting (5000 words, 6 subjects): ~17 + 12 + 3 = **32 calls**
## S3 Operations Summary
### Source Bucket: `DAILYCO_STORAGE_AWS_BUCKET_NAME`
Daily.co uploads raw-tracks recordings here.
| Operation | Key Pattern | When |
|-----------|-------------|------|
| **READ** (presign) | `{domain}/{room_name}/{ts}/{participant_id}-cam-audio-{ts}.webm` | Track acquisition |
| **DELETE** | Same as above | Consent denied cleanup |
### Transcript Storage Bucket: `TRANSCRIPT_STORAGE_AWS_BUCKET_NAME`
Reflector's own storage.
| Operation | Key Pattern | When |
|-----------|-------------|------|
| **PUT** | `file_pipeline/{transcript_id}/tracks/padded_{idx}.webm` | After track padding |
| **READ** (presign) | Same | For GPU transcription |
| **DELETE** | Same | After transcription complete |
| **PUT** | `{transcript_id}/audio.mp3` | After mixdown |
| **DELETE** | Same | Consent denied cleanup |
## Database Operations
### Tables Written
| Table | Operation | When |
|-------|-----------|------|
| `recording` | INSERT | Initialization |
| `transcript` | INSERT | Initialization |
| `transcript` | UPDATE (participants) | After Daily API participant fetch |
| `transcript` | UPDATE (status, events, duration, topics, title, summaries, etc.) | Throughout pipeline |
### Transcript Update Sequence
```
1. INSERT: id, name, status='idle', source_kind='room', user_id, recording_id, room_id, meeting_id
2. UPDATE: participants[] (speaker index → participant name mapping)
3. UPDATE: status='processing', events+=[{event:'STATUS', data:{value:'processing'}}]
4. UPDATE: duration=X, events+=[{event:'DURATION', data:{duration:X}}]
5. UPDATE: audio_location='storage'
6. UPDATE: events+=[{event:'WAVEFORM', data:{waveform:[...]}}]
7. UPDATE: events+=[{event:'TRANSCRIPT', data:{text, translation}}]
8. UPDATE: topics[]+=topic, events+=[{event:'TOPIC'}] -- repeated per chunk
9. UPDATE: title=X, events+=[{event:'FINAL_TITLE'}]
10. UPDATE: long_summary=X, events+=[{event:'FINAL_LONG_SUMMARY'}]
11. UPDATE: short_summary=X, events+=[{event:'FINAL_SHORT_SUMMARY'}]
12. UPDATE: status='ended', events+=[{event:'STATUS', data:{value:'ended'}}]
13. UPDATE: zulip_message_id=X -- if Zulip enabled
14. UPDATE: audio_deleted=true -- if consent denied
```
## WebSocket Events
All broadcast to room `ts:{transcript_id}`:
| Event | Payload | Trigger |
|-------|---------|---------|
| STATUS | `{value: "processing"\|"ended"\|"error"}` | Status transitions |
| DURATION | `{duration: float}` | After audio processing |
| WAVEFORM | `{waveform: float[]}` | After waveform generation |
| TRANSCRIPT | `{text: string, translation: string\|null}` | After transcription merge |
| TOPIC | `{id, title, summary, timestamp, duration, transcript, words}` | Per topic detected |
| FINAL_TITLE | `{title: string}` | After LLM title generation |
| FINAL_LONG_SUMMARY | `{long_summary: string}` | After LLM summary |
| FINAL_SHORT_SUMMARY | `{short_summary: string}` | After LLM recap |
User-room broadcasts to `user:{user_id}`:
- `TRANSCRIPT_STATUS`
- `TRANSCRIPT_FINAL_TITLE`
- `TRANSCRIPT_DURATION`
Documentation coming soon.

421
server/docs/daily_pipeline.md Normal file
View File

@@ -0,0 +1,421 @@
# Daily.co pipeline
This document details every external call, storage operation, and database write that occurs when a new Daily.co recording is discovered.
It includes a bunch of common logic that other pipelines use, therefore not everything is Daily-oriented.
**The doc was generated at 12.12.2025 and things may have changed since.**
## Trigger
Two entry points, both converging to the same handler:
1. **Webhook**: Daily.co sends `POST /v1/daily/webhook` with `recording.ready-to-download`
2. **Polling**: `GET /recordings` (paginated, max 100/call) → filter new → convert to same payload format
Both produce `RecordingReadyPayload` and call `handleRecordingReady(payload)`.
```
┌─────────────────┐ ┌──────────────────────────┐
│ Daily Webhook │────▶│ RecordingReadyPayload │
│ (push) │ │ {room_name, recording_id│
└─────────────────┘ │ tracks[], ...} │
└────────────┬─────────────┘
┌─────────────────┐ │
│ GET /recordings│ ▼
│ (poll) │────▶ convert ──▶ handleRecordingReady()
└─────────────────┘ │
▼
┌────────────────────────┐
│ process_multitrack_ │
│ recording pipeline │
└────────────────────────┘
```
**Polling API**: `GET https://api.daily.co/v1/recordings`
- Pagination: `limit` (max 100), `starting_after`, `ending_before`
- Rate limit: ~2 req/sec
- Response: `{total_count, data: Recording[]}`
```mermaid
flowchart TB
subgraph Trigger["1. Recording Discovery - Daily.co Webhook"]
DAILY_WEBHOOK["Daily.co sends POST /v1/daily/webhook<br/>type: recording.ready-to-download"]
VERIFY["Verify X-Webhook-Signature (HMAC)"]
PARSE["Parse DailyWebhookEvent<br/>Extract tracks[], room_name, recording_id"]
FILTER["Filter audio tracks only<br/>track_keys = [t.s3Key for t in tracks if t.type == 'audio']"]
DISPATCH["process_multitrack_recording.delay()"]
DAILY_WEBHOOK --> VERIFY --> PARSE --> FILTER --> DISPATCH
end
subgraph Init["2. Recording Initialization"]
FETCH_MEETING[DB READ: meetings_controller.get_by_room_name]
FETCH_ROOM[DB READ: rooms_controller.get_by_name]
DAILY_API_REC[Daily API: GET /recordings/recording_id]
DAILY_API_PART[Daily API: GET /meetings/mtgSessionId/participants]
CREATE_RECORDING[DB WRITE: recordings_controller.create]
CREATE_TRANSCRIPT[DB WRITE: transcripts_controller.add]
MAP_PARTICIPANTS[DB WRITE: transcript.participants upsert]
end
subgraph Pipeline["3. Processing Pipeline"]
direction TB
PAD[Track Padding & Mixdown]
TRANSCRIBE[GPU: Transcription per track]
TOPICS[LLM: Topic Detection]
TITLE[LLM: Title Generation]
SUMMARY[LLM: Summary Generation]
end
subgraph Storage["4. S3 Operations"]
S3_PRESIGN[S3: generate_presigned_url for tracks]
S3_UPLOAD_PADDED[S3 UPLOAD: padded tracks temp]
S3_UPLOAD_MP3[S3 UPLOAD: audio.mp3]
S3_DELETE_TEMP[S3 DELETE: cleanup temp files]
end
subgraph PostProcess["5. Post-Processing"]
CONSENT[Consent check & cleanup]
ZULIP[Zulip: send/update message]
WEBHOOK_OUT[Webhook: POST to room.webhook_url]
end
Trigger --> Init --> Pipeline
Pipeline --> Storage
Pipeline --> PostProcess
```
## Detailed Sequence: Daily.co Multitrack Recording
```mermaid
sequenceDiagram
participant DailyCo as Daily.co
participant API as FastAPI /v1/daily/webhook
participant Worker as Celery Worker
participant DB as PostgreSQL
participant DailyAPI as Daily.co REST API
participant S3 as AWS S3
participant GPU as Modal.com GPU
participant LLM as LLM Service
participant WS as WebSocket
participant Zulip as Zulip
participant ExtWH as External Webhook
Note over DailyCo,API: Phase 0: Webhook Receipt
DailyCo->>API: POST /v1/daily/webhook
Note right of DailyCo: X-Webhook-Signature, X-Webhook-Timestamp
API->>API: verify_webhook_signature()
API->>API: Extract audio track s3Keys from payload.tracks[]
API->>Worker: process_multitrack_recording.delay()
API-->>DailyCo: 200 OK
Note over Worker,DailyAPI: Phase 1: Recording Initialization
Worker->>DB: SELECT meeting WHERE room_name=?
Worker->>DB: SELECT room WHERE name=?
Worker->>DailyAPI: GET /recordings/{recording_id}
DailyAPI-->>Worker: {mtgSessionId, ...}
Worker->>DailyAPI: GET /meetings/{mtgSessionId}/participants
DailyAPI-->>Worker: [{participant_id, user_name}, ...]
Worker->>DB: INSERT INTO recording
Worker->>DB: INSERT INTO transcript (status='idle')
loop For each track_key (parse participant_id from filename)
Worker->>DB: UPSERT transcript.participants[speaker=idx, name=X]
end
Note over Worker,S3: Phase 2: Track Padding
Worker->>DB: UPDATE transcript SET status='processing'
Worker->>WS: broadcast STATUS='processing'
loop For each track in track_keys (N tracks)
Worker->>S3: generate_presigned_url(track_key, DAILYCO_BUCKET)
S3-->>Worker: presigned_url (2hr)
Note over Worker: PyAV: read WebM, extract start_time
Note over Worker: PyAV: adelay filter (pad silence)
Worker->>S3: PUT file_pipeline/{id}/tracks/padded_{idx}.webm
Worker->>S3: generate_presigned_url(padded_{idx}.webm)
end
Note over Worker,S3: Phase 3: Audio Mixdown
Note over Worker: PyAV: amix filter → stereo MP3
Worker->>DB: UPDATE transcript SET duration=X
Worker->>WS: broadcast DURATION
Worker->>S3: PUT {transcript_id}/audio.mp3
Worker->>DB: UPDATE transcript SET audio_location='storage'
Note over Worker: Phase 4: Waveform
Note over Worker: Generate peaks from MP3
Worker->>DB: UPDATE events+=WAVEFORM
Worker->>WS: broadcast WAVEFORM
Note over Worker,GPU: Phase 5: Transcription (N GPU calls)
loop For each padded track URL (N tracks)
Worker->>GPU: POST /v1/audio/transcriptions-from-url
Note right of GPU: {audio_file_url, language, batch:true}
GPU-->>Worker: {words: [{word, start, end}, ...]}
Note over Worker: Assign speaker=track_idx to words
end
Note over Worker: Merge all words, sort by start time
Worker->>DB: UPDATE events+=TRANSCRIPT
Worker->>WS: broadcast TRANSCRIPT
Note over Worker,S3: Cleanup temp files
loop For each padded file
Worker->>S3: DELETE padded_{idx}.webm
end
Note over Worker,LLM: Phase 6: Topic Detection (C LLM calls)
Note over Worker: C = ceil(total_words / 300)
loop For each 300-word chunk (C chunks)
Worker->>LLM: TOPIC_PROMPT + words[i:i+300]
Note right of LLM: "Extract main topic title + 2-sentence summary"
LLM-->>Worker: TitleSummary{title, summary}
Worker->>DB: UPSERT topics[]
Worker->>DB: UPDATE events+=TOPIC
Worker->>WS: broadcast TOPIC
end
Note over Worker,LLM: Phase 7a: Title Generation (1 LLM call)
Note over Worker: Input: all TitleSummary[].title joined
Worker->>LLM: TITLE_PROMPT
Note right of LLM: "Generate concise title from topic titles"
LLM-->>Worker: "Meeting Title"
Worker->>DB: UPDATE transcript SET title=X
Worker->>DB: UPDATE events+=FINAL_TITLE
Worker->>WS: broadcast FINAL_TITLE
Note over Worker,LLM: Phase 7b: Summary Generation (2+2M LLM calls)
Note over Worker: Reconstruct full transcript from TitleSummary[].transcript
opt If participants unknown
Worker->>LLM: PARTICIPANTS_PROMPT
LLM-->>Worker: ParticipantsResponse
end
Worker->>LLM: SUBJECTS_PROMPT (call #1)
Note right of LLM: "Main high-level topics? Max 6"
LLM-->>Worker: SubjectsResponse{subjects: ["A", "B", ...]}
loop For each subject (M subjects, max 6)
Worker->>LLM: DETAILED_SUBJECT_PROMPT (call #2..#1+M)
Note right of LLM: "Info about 'A': decisions, actions, deadlines"
LLM-->>Worker: detailed_response (discarded after next call)
Worker->>LLM: PARAGRAPH_SUMMARY_PROMPT (call #2+M..#1+2M)
Note right of LLM: "Summarize in 1 paragraph"
LLM-->>Worker: paragraph → summaries[]
end
Worker->>LLM: RECAP_PROMPT (call #2+2M)
Note right of LLM: "High-level quick recap, 1 paragraph"
LLM-->>Worker: recap
Note over Worker: long_summary = "# Quick recap\n{recap}\n# Summary\n**A**\n{para1}..."
Note over Worker: short_summary = recap only
Worker->>DB: UPDATE long_summary, short_summary
Worker->>DB: UPDATE events+=FINAL_LONG_SUMMARY
Worker->>WS: broadcast FINAL_LONG_SUMMARY
Worker->>DB: UPDATE events+=FINAL_SHORT_SUMMARY
Worker->>WS: broadcast FINAL_SHORT_SUMMARY
Note over Worker,DB: Phase 8: Finalize
Worker->>DB: UPDATE transcript SET status='ended'
Worker->>DB: UPDATE events+=STATUS
Worker->>WS: broadcast STATUS='ended'
Note over Worker,ExtWH: Phase 9: Post-Processing Chain
Worker->>DB: SELECT meeting_consent WHERE meeting_id=?
alt Any consent denied
Worker->>S3: DELETE tracks from DAILYCO_BUCKET
Worker->>S3: DELETE audio.mp3 from TRANSCRIPT_BUCKET
Worker->>DB: UPDATE transcript SET audio_deleted=true
end
opt Room has zulip_auto_post=true
alt Existing zulip_message_id
Worker->>Zulip: PATCH /api/v1/messages/{id}
else New
Worker->>Zulip: POST /api/v1/messages
Zulip-->>Worker: {id}
Worker->>DB: UPDATE transcript SET zulip_message_id=X
end
end
opt Room has webhook_url
Worker->>ExtWH: POST {webhook_url}
Note right of ExtWH: X-Webhook-Signature: HMAC-SHA256
Note right of ExtWH: Body: {transcript_id, room_id, ...}
end
```
## Title & Summary Generation Data Flow
```mermaid
flowchart TB
subgraph Input["Input: TitleSummary[] from Topic Detection"]
TS1["TitleSummary 1<br/>title: 'Q1 Budget'<br/>transcript: words[0:300]"]
TS2["TitleSummary 2<br/>title: 'Product Launch'<br/>transcript: words[300:600]"]
TS3["TitleSummary N..."]
end
subgraph TitleGen["Title Generation"]
T1["Extract .title from each TitleSummary"]
T2["Concatenate: '- Q1 Budget\n- Product Launch\n...'"]
T3["LLM: TITLE_PROMPT\n'Generate concise title from topic titles'"]
T4["Output: FinalTitle"]
T1 --> T2 --> T3 --> T4
end
subgraph SummaryGen["Summary Generation"]
direction TB
subgraph Reconstruct["1. Reconstruct Full Transcript"]
S1["For each TitleSummary.transcript.as_segments()"]
S2["Map speaker ID → name"]
S3["Build: 'Alice: hello\nBob: hi\n...'"]
S1 --> S2 --> S3
end
subgraph Subjects["2. Extract Subjects - LLM call #1"]
S4["LLM: SUBJECTS_PROMPT\n'Main high-level topics? Max 6'"]
S5["subjects[] = ['Budget Review', ...]"]
S4 --> S5
end
subgraph DetailedSum["3. Per-Subject Summary - LLM calls #2 to #(1+2M)"]
S6["For each subject:"]
S7["LLM: DETAILED_SUBJECT_PROMPT\n'Info about subject: decisions, actions...'"]
S8["detailed_response - NOT STORED"]
S9["LLM: PARAGRAPH_SUMMARY_PROMPT\n'Summarize in 1 paragraph'"]
S10["paragraph → summaries[]"]
S6 --> S7 --> S8 --> S9 --> S10
end
subgraph Recap["4. Generate Recap - LLM call #(2+2M)"]
S11["Concatenate paragraph summaries"]
S12["LLM: RECAP_PROMPT\n'High-level recap, 1 paragraph'"]
S13["recap"]
S11 --> S12 --> S13
end
subgraph Output["5. Output"]
S14["long_summary = markdown:\n# Quick recap\n[recap]\n# Summary\n**Subject 1**\n[para1]..."]
S15["short_summary = recap only"]
S14 --> S15
end
Reconstruct --> Subjects --> DetailedSum --> Recap --> Output
end
Input --> TitleGen
Input --> SummaryGen
```
### topics[] vs subjects[]
| | topics[] | subjects[] |
|-|----------|------------|
| **Source** | 300-word chunk splitting | LLM extraction from full text |
| **Count** | Variable (words / 300) | Max 6 |
| **Purpose** | Timeline segmentation | Summary structure |
| **Has timestamp?** | Yes | No |
## External API Calls Summary
### 1. Daily.co REST API (called during initialization)
| Endpoint | Method | When | Purpose |
|----------|--------|------|---------|
| `GET /recordings/{recording_id}` | GET | After webhook | Get mtgSessionId for participant lookup |
| `GET /meetings/{mtgSessionId}/participants` | GET | After above | Map participant_id → user_name |
### 2. GPU Service (Modal.com or Self-Hosted)
| Endpoint | Method | Count | Request |
|----------|--------|-------|---------|
| `{TRANSCRIPT_URL}/v1/audio/transcriptions-from-url` | POST | **N** (N = num tracks) | `{audio_file_url, language, batch: true}` |
**Note**: Diarization is NOT called for multitrack - speaker identification comes from separate tracks.
### 3. LLM Service (OpenAI-compatible via LlamaIndex)
| Phase | Operation | Input | LLM Calls | Output |
|-------|-----------|-------|-----------|--------|
| Topic Detection | TOPIC_PROMPT per 300-word chunk | words[i:i+300] | **C** = ceil(words/300) | TitleSummary{title, summary, timestamp} |
| Title Generation | TITLE_PROMPT | All topic titles joined | **1** | FinalTitle |
| Participant ID | PARTICIPANTS_PROMPT | Full transcript | **0-1** (skipped if known) | ParticipantsResponse |
| Subject Extraction | SUBJECTS_PROMPT | Full transcript | **1** | SubjectsResponse{subjects[]} |
| Subject Detail | DETAILED_SUBJECT_PROMPT | Full transcript + subject name | **M** (M = subjects, max 6) | detailed text (discarded) |
| Subject Paragraph | PARAGRAPH_SUMMARY_PROMPT | Detailed text | **M** | paragraph text → summaries[] |
| Recap | RECAP_PROMPT | All paragraph summaries | **1** | recap text |
**Total LLM calls**: C + 2M + 3 (+ 1 if participants unknown)
- Short meeting (1000 words, 3 subjects): ~4 + 6 + 3 = **13 calls**
- Long meeting (5000 words, 6 subjects): ~17 + 12 + 3 = **32 calls**
## S3 Operations Summary
### Source Bucket: `DAILYCO_STORAGE_AWS_BUCKET_NAME`
Daily.co uploads raw-tracks recordings here.
| Operation | Key Pattern | When |
|-----------|-------------|------|
| **READ** (presign) | `{domain}/{room_name}/{ts}/{participant_id}-cam-audio-{ts}.webm` | Track acquisition |
| **DELETE** | Same as above | Consent denied cleanup |
### Transcript Storage Bucket: `TRANSCRIPT_STORAGE_AWS_BUCKET_NAME`
Reflector's own storage.
| Operation | Key Pattern | When |
|-----------|-------------|------|
| **PUT** | `file_pipeline/{transcript_id}/tracks/padded_{idx}.webm` | After track padding |
| **READ** (presign) | Same | For GPU transcription |
| **DELETE** | Same | After transcription complete |
| **PUT** | `{transcript_id}/audio.mp3` | After mixdown |
| **DELETE** | Same | Consent denied cleanup |
## Database Operations
### Tables Written
| Table | Operation | When |
|-------|-----------|------|
| `recording` | INSERT | Initialization |
| `transcript` | INSERT | Initialization |
| `transcript` | UPDATE (participants) | After Daily API participant fetch |
| `transcript` | UPDATE (status, events, duration, topics, title, summaries, etc.) | Throughout pipeline |
### Transcript Update Sequence
```
1. INSERT: id, name, status='idle', source_kind='room', user_id, recording_id, room_id, meeting_id
2. UPDATE: participants[] (speaker index → participant name mapping)
3. UPDATE: status='processing', events+=[{event:'STATUS', data:{value:'processing'}}]
4. UPDATE: duration=X, events+=[{event:'DURATION', data:{duration:X}}]
5. UPDATE: audio_location='storage'
6. UPDATE: events+=[{event:'WAVEFORM', data:{waveform:[...]}}]
7. UPDATE: events+=[{event:'TRANSCRIPT', data:{text, translation}}]
8. UPDATE: topics[]+=topic, events+=[{event:'TOPIC'}] -- repeated per chunk
9. UPDATE: title=X, events+=[{event:'FINAL_TITLE'}]
10. UPDATE: long_summary=X, events+=[{event:'FINAL_LONG_SUMMARY'}]
11. UPDATE: short_summary=X, events+=[{event:'FINAL_SHORT_SUMMARY'}]
12. UPDATE: status='ended', events+=[{event:'STATUS', data:{value:'ended'}}]
13. UPDATE: zulip_message_id=X -- if Zulip enabled
14. UPDATE: audio_deleted=true -- if consent denied
```
## WebSocket Events
All broadcast to room `ts:{transcript_id}`:
| Event | Payload | Trigger |
|-------|---------|---------|
| STATUS | `{value: "processing"\|"ended"\|"error"}` | Status transitions |
| DURATION | `{duration: float}` | After audio processing |
| WAVEFORM | `{waveform: float[]}` | After waveform generation |
| TRANSCRIPT | `{text: string, translation: string\|null}` | After transcription merge |
| TOPIC | `{id, title, summary, timestamp, duration, transcript, words}` | Per topic detected |
| FINAL_TITLE | `{title: string}` | After LLM title generation |
| FINAL_LONG_SUMMARY | `{long_summary: string}` | After LLM summary |
| FINAL_SHORT_SUMMARY | `{short_summary: string}` | After LLM recap |
User-room broadcasts to `user:{user_id}`:
- `TRANSCRIPT_STATUS`
- `TRANSCRIPT_FINAL_TITLE`
- `TRANSCRIPT_DURATION`