Haystack

In this catalog it is the RAG option when you need to ground model answers in retrieved company documents. Keep Haystack on the observe/propose side of the loop: useful signals and drafts, with human or policy authority still required before anything irreversible runs.

Ingest parsers chunk and embed content into a store; at query time the pipeline retrieves, optionally reranks, and feeds context into the LLM.

1. Ingest a sample document through the product’s parser/index path.
2. Embed and store chunks in the configured vector backend.
3. Ask a grounded question; confirm citations/hits come from the fixture.
4. Wire the agent host to retrieve before it proposes actions.

Related: llamaindex · qdrant · docling · topics/04-data · topics/35-context-engineering

• Use when: Haystack (deepset) is a mature open framework for NLP/RAG pipelines: document stores, retrievers, rankers, and production connectors matches the bottleneck and you specifically need to ground model answers in retrieved company documents.
• Skip when: another pack already covers the need with less ops surface, or Haystack would become a second place people retype truth.
• Prefer instead: compare LlamaIndex, Qdrant, Docling before locking the cell.

• Ops: confirm license, hosting, and release cadence for Haystack (deepset-ai/haystack) before relying on it in a live corridor.
• Integration: bad credentials, scopes, or schema drift produce empty retrievals or runaway tool/model spend without an obvious UI alarm.
• Scope: RAG components support agents; they are not the company ledger or the approval authority.