01
Garbage in still means garbage out
RAG quality is mostly data quality. Stale docs, duplicate policies, and missing product updates will surface as confident wrong answers.
Treat your knowledge base like a product surface: owners, freshness SLAs, and a clear source of truth.
02
Chunk for questions, not for storage
Arbitrary token splits break meaning. Chunk around headings, procedures, and FAQ units so retrieval returns complete thoughts.
Then tune overlap carefully — enough context to answer, not so much that ranking becomes noise.
03
Citations are a UX feature
Showing sources is not just compliance theater. Citations let users verify answers and teach the system where trust breaks.
If an answer cannot point to a source, the UI should say so — and offer a path to ask a human.



