RAG quality starts with knowledge design

When an assistant answers from company documents, the model is only one part of the system. The result also depends on which material is available, how it is divided, whether the user is allowed to see it and how relevance is judged. If source content is stale or contradictory, retrieval can confidently surface the wrong version. Prepare the knowledge before indexing Inventory the sources and identify an owner, review date, audience and status for each one. Remove duplicate drafts, separate policy from discussion and mark superseded instructions clearly. A search index should not turn every shared file into approved company guidance. Start with a small, representative collection that has a reason to be trusted. Preserve meaning in chunks Split material around meaningful sections rather than arbitrary character counts alone. Keep headings, definitions, dates and relevant relationships attached to the passage they explain. Store source identifiers and access metadata so the system can cite where an answer came from and enforce the same permissions as the underlying content. Evaluate retrieval and answers separately Build a test set from real questions, including ambiguous and unanswerable examples. First ask whether the right source passages appear near the top. Then assess whether the answer is supported by those passages, whether uncertainty is communicated and whether a citation leads to useful evidence. This separates a retrieval problem from a generation problem. Plan for no answer A dependable assistant must be able to say that the available sources do not establish an answer. It should offer a safe route to a human or authoritative system, especially for policy, security and customer commitments. Monitor failed retrievals, user corrections, stale sources and permission denials. Retrieval quality is an ongoing content and operations practice, not a one-time model setting.