RAG Pipelines That Do Not Hallucinate Your Docs

Chunking, citations, and evals we use when a client wants an LLM that actually stays on their knowledge base.

Retrieval is the product

A chatbot that “knows your docs” fails in the retrieval layer long before the model does. We start with source of truth, chunk sizes that match how people ask questions, and a citation UI that makes misses obvious.

We never ship a RAG demo without an eval set of 40–80 real questions from support tickets. If the answer cannot point at a chunk, it is a refusal, not a guess.

Stack we reach for

Embeddings + hybrid search (keyword + vector) for SKUs and policy language.
Reranking on the top 20 hits before the prompt is built.
Structured output so the UI can render sources next to the answer.

Takeaway

If you cannot audit why the model said something, you do not have a knowledge assistant. You have a very confident intern.