RAG retriever — vector search endpoint returning top-K chunks for a query

[Punkte]

What to build

A retriever service and supporting endpoint that embeds an incoming query and returns the top-K most semantically similar conversation chunks from pgvector. Used by the chat engine to inject factual context into each prompt.

Acceptance criteria

• retrieveContext(conversationId, query, { topK, scoreThreshold }) returns ranked chunks with cosine similarity scores
• Uses pgvector <=> cosine distance operator
• Default: top 5 chunks, minimum score 0.3
• GET /conversations/:id/context?q=... endpoint exposes retriever for testing
• Returns chunks with metadata (date range, participants, score)
• Verified: semantically relevant chunks rank higher than unrelated ones on a test query

Blocked by

• Chunking + embedding pipeline — temporal sliding window → pgvector #4