Prevent LLM Hallucinations Consulting

Put the right evidence in context with hybrid retrieval, forbid answers when retrieval is empty, require citations to spans, and measure faithfulness on a golden set in CI. Prompting “don’t make things up” does not hold. We install those controls in the client cloud in four weeks standard, with per-user ACL at query time so the model cannot quote documents the user should not see.

Usually the retriever returned the wrong chunk, the prompt allowed ungrounded claims, or empty retrieval still generated a paragraph. Sometimes citations are decorative. The fix is better hybrid retrieval, a hard refuse-when-empty path, citation checks, and CI evals — not a larger model. Week 1 of this engagement is an audit that names which of those four you actually have.

For facts that live in your documents, no — not reliably. Closed-book models will invent. Fine-tuning does not keep weekly policy changes current. If the task is style or format, grounding is the wrong tool. Discovery separates fact questions (retrieval plus refusal) from skill questions (prompts, evals, maybe fine-tuning). We will not sell retrieval for a tone problem.

If hybrid retrieval returns no ACL-allowed chunk, the system says it does not know and does not answer from model memory. That rule is tested in CI with dedicated cases. Teams remove it under demo pressure; we put it back. Citations without this rule still allow fluent invention when the index misses.

A standard control set is four weeks: audit, wiring, golden set, shadow-mode, handover, then 30 days on-call. Financial services typically 8–12 weeks. Healthcare typically 10–14. Ecommerce typically 6–10. If your stack cannot retrieve or trace, that work is in scope and we write it down before week 2.

Not as the default. There is no shared training. Client-owned fine-tunes are optional when evals show a skill gap retrieval cannot fix. Truthfulness for enterprise facts is a retrieval and refusal problem. You pay inference. We are model-agnostic. Work runs in your cloud. You own the IP.

Claim-level faithfulness against retrieved sources on a golden set, plus live sampling. We do not quote a universal “hallucination percent” as a marketing number. CI tracks the delta versus the last pin. Empty-retrieval cases are scored separately so a chatty model cannot look good by answering everything.

Fixed-scope implementation fee plus platform fee, quoted in writing. You pay inference. You own prompts, retrieval config, evals, traces, and runbooks in the client cloud. No shared training. Thirty days on-call included. Further red-team cases after that window are a new fixed scope.