Spike: Haystack 3 vs in-tree Librarian for chat RAG

[northdpole]

Task

Time-boxed research spike (not an adoption PR): does Haystack 3 (Apache-2.0) buy anything OpenCRE does not already have in Librarian C.1/C.2 + title_boost, for chat RAG after #1075?

Default bias from maintainer research: do not replace /completion with a Haystack app. This spike exists to kill or narrow that option with evidence.

Success criteria (all must pass)

• Written verdict in this issue: adopt / don’t adopt / steal one component
• Compare on OpenCRE constraints, not generic RAG blogs:
  • Prod Heroku slug: no torch / sentence-transformers in requirements.txt
  • pgvector already in Postgres; CRE + Standard embedding tables
  • Must emit CRE id + standard URL citations
  • Existing PromptHandler + Google login + spend limits
• If “steal one component”: name the class (e.g. hybrid BM25+dense, eval pipeline) and an estimate of lines vs wrapping PgVectorRetriever
• Kill criteria: if Haystack is only a wrapper around “embed → pgvector top-k → rerank → prompt”, don’t adopt
• License/ops: Apache-2.0 is fine; note extra services (Elastic/OpenSearch) if any pipeline we care about requires them
• No production dependency added in this spike. A spike branch/doc is OK; do not merge Haystack onto main from this ticket

Context

• In-tree: application/utils/librarian/ (retriever, CE rerank, factory)
• Chat: application/prompt_client/prompt_client.py generate_text
• Related: Chat: retrieve top-k via Librarian seams (implements #962) #1075 (implementation), Spike: RAGAS (or equivalent) eval for cited OpenCRE Chat #1078 (RAGAS spike), MyOpenCRE mapping: reuse Module C Librarian retriever/reranker (RFC TODO 3) #965 (MyOpenCRE)

Constraints

• Time-box: 1–2 days of reading + a throwaway notebook/script against a tiny local fixture, not a platform rewrite
• Out of scope: Dify, RAGFlow, Onyx, Open Notebook, LlamaIndex-as-the-chat-app (LlamaIndex stays harvester-dev only)