Retrieval-augmented generation was supposed to solve hallucination. In practice, naive RAG often just gives hallucination better props — the model still improvises, now with a stack of documents as set dressing.
Retrieval is a search problem
Good RAG starts with taking retrieval seriously: hybrid semantic and keyword search, chunking that respects document structure, freshness weighting so yesterday's policy beats last year's, and permission-aware filtering so people only ever see what they're allowed to see. If the right passage isn't retrieved, no prompt can save the answer.
Answers with receipts
Then comes the part most implementations skip: verification. Krutakam Kosh checks that each claim in a generated answer is actually supported by a retrieved passage, and every answer ships with citations that link straight to the source. When support is weak, the system says "I don't know" — which, in an enterprise, is a feature.
The result is boring in the best way: answers people can check, and therefore answers people actually rely on.