
Retrieval-augmented generation is how many product teams ground answers in their own content. The hard parts are rarely the embedding model — they are freshness, access control, and knowing when the system should say “I do not know.”
Treat documents as a product
If the source set is a folder nobody owns, the agent will confidently cite last year’s policy. Assign ownership, version docs, and refresh the index on a schedule you can defend.
Permissions travel with the chunk
A vector store without ACLs will leak. Filter retrieval by what the signed-in user may see, and test that filter the same way you test any other authorization path.
Refuse when evidence is thin
Better a clear “not in our docs” than a fluent guess. Thresholds on retrieval score, citation of sources, and a path for a human to correct the corpus keep trust intact.
Grounded answers are a data and permissions problem as much as an AI problem. We wire RAG the same way we ship any other production feature: owned sources, tests, and monitoring.

