feat: Whisper audio transcription pipeline for audio documents

[nv78]

Summary

When a user uploads an audio file (MP3, M4A, WAV, OGG, FLAC, WebM, etc.) as a document, the new pipeline transcribes it via OpenAI Whisper and feeds the transcript into the existing RAG chunking + embedding pipeline. The audio becomes fully searchable — ask "what did the speaker say about X?" and it works.

New file: backend/services/audio_service.py

• transcribe_audio(bytes, filename, language?) — main entry point
• Calls openai.OpenAI().audio.transcriptions.create() with whisper-1 and response_format="verbose_json" to capture detected language and duration
• Prepends a metadata header to the transcript so it surfaces in search:

  [Audio transcript — language: en, duration: 3m 42s]
  
  The speaker begins by discussing...
  

• Passes a prompt hint to Whisper to improve punctuation and paragraph breaks
• Validates file size against MAX_AUDIO_BYTES (default 25 MB — Whisper's hard limit)
• Validates extension against the set Whisper actually supports: flac, m4a, mp3, mp4, mpeg, mpga, oga, ogg, wav, webm
• Respects ENABLE_MULTIMODAL flag
• Never raises — returns a placeholder string on failure so the document record is always created
• language param accepts BCP-47 codes ("en", "es", etc.) for better accuracy when known

Updated: backend/api_endpoints/documents/handler.py

New elif category == "audio" branch in IngestDocumentsHandler:

1. Reads raw bytes from the upload
2. Calls transcribe_audio(audio_bytes, filename=filename)
3. Stores transcript as document_text (media_type='audio', mime_type)
4. Enqueues chunk_document_fn.remote() for embedding

Video remains a binary-only stub — to be wired in PR 4.

Test plan

• Upload an MP3 → document record created with media_type='audio', transcript in document_text
• Chunks appear in chunks table after Ray task completes
• Ask a question about spoken content → RAG retrieves relevant transcript chunk, LLM answers correctly
• Upload a 26 MB audio file → size-limit placeholder stored, no API call made
• Upload a .wav file → transcription works
• Upload a .pdf → text path unaffected (no regression)
• Set ENABLE_MULTIMODAL=false → placeholder stored, no Whisper call

Depends on: claude/image-document-analysis → claude/add-multimodal-support-QBQca

https://claude.ai/code/session_01C9mHttiQ4ZAaBbQecVV7uu