How do I stop an AI assistant making things up about my business?

Hallucination is largely a question of what happens when the right information is absent. A model asked something its context does not cover will produce a confident answer anyway, because producing text is what it does. The fix is architectural rather than a matter of prompting more firmly: retrieve first, score what came back, and treat a low score as a reason to refuse.

Retrieval quality matters as much as the floor. Semantic search alone misses exact tokens — a policy number, a plan name, a registration — because two strings can mean similar things and share no vocabulary. Keyword search alone misses paraphrase. Running both and fusing the results, which is what reciprocal rank fusion does, catches considerably more than either. RegenticFlow runs vector search and BM25 on every question and fuses them.

Citations do a second job beyond reassurance: they make wrong answers findable. An uncited answer that misstates your refund window looks exactly like a correct one. An answer with the document named underneath can be checked in seconds, by you or by the customer.

The remaining risk is silent drift. Retrieval that worked in January can stop reaching the right document after you upload forty more, and nothing announces it. Saved checks re-run known questions when your knowledge changes and email you when a document stops reaching the answer it used to support.

• None of this makes an answer guaranteed correct. It narrows the failure mode from invention to misreading, which is a smaller problem, not no problem.
• Checks of that kind verify retrieval — whether the right source still reaches the answer. They do not grade the wording of the reply. Doing that would need a second model judging every run.
• An assistant is only as good as its documents. A five-page policy PDF makes a five-page-good assistant, and no amount of retrieval engineering changes that.

• Where the limits are, in full
• vs Chatbase