Vectorise

Second stage of the core pipeline: turn stored text (and voice transcripts) into chunked vector embeddings in Neo4j.

Voice path

After ingest creates a Voice node:

/api/story/voice-vectorise?voiceId=…&mode=chain
  → transcribe (Whisper, ≤180s inline)
  → triggerResolve (if channel has resolve route)
  → vectorise (chunk + embed)

Long voice (>180s): marked deferred_long, handled by Railway transcribe service (.transcribe/), then vectorised on return.

Cron routes: /api/story/voice-vectorise, /api/story/transcribe sweep pending and failed work.

Other sources

Same embedding utilities (services/vectorise/shared/):

All write to Neo4j vector indexes for downstream apps like evaluate (/api/rag).

Storage

Embeddings and chunks live on Neo4j nodes — not a separate vector DB. Requires OPENAI_API_KEY for Whisper and embeddings.

Failed transcribe/vectorise stages queue to timelining::transcribe::failed.

Code

src/services/vectorise/ — voice/, page/, resource/, shared/