Chunking + embedding pipeline — temporal sliding window → pgvector

[Punkte]

What to build

After upload, automatically split the conversation into overlapping temporal chunks and generate embeddings for each via OpenAI text-embedding-3-small, storing vectors in pgvector. This pipeline runs as part of the post-upload processing flow.

Acceptance criteria

• chunkConversation() splits messages using temporal sliding window (max 30 messages, 60-min gap cut, 5-message overlap)
• Each chunk formatted as readable text with timestamps and sender names
• Embeddings generated via text-embedding-3-small in batches of 100
• Chunks + embeddings stored in chunks table with metadata (date range, participants)
• Conversation status updated to 'ready' after pipeline completes
• Smoke test: query "do you remember the holidays?" returns relevant chunks

Blocked by

• Upload endpoint — POST /upload detects format, parses, stores conversation + messages #3