What RAG is for
Retrieval-augmented generation lets Claude answer from your wikis, tickets, and manuals instead of guessing. For an enterprise chatbot, retrieval quality is the product.
Pipeline stages
Ingest, chunk with overlap, embed, index with metadata, retrieve, optional re-rank, then generate with citations. Weak chunking shows up as missed SKUs and hallucinated policy.
Eval harness
Measure hit rate and faithfulness on golden questions whenever you change models or chunk sizes. Block deploys that fail the suite. That is how InferentCloud keeps Knowledge Assistant and support chatbots trustworthy.
Access control
Filter chunks by tenant and role before they reach Claude. A great answer from the wrong document is still a failure.
Summary
RAG plus a Claude agentic harness (tools, citations, evals) is the default for operator-grade knowledge chat.