Running a Hallucination Audit on Your Prompts
You cannot eliminate hallucination. You can find where your prompts invite it and wall those spots off. A repeatable audit process.
Hallucination is not a bug you fix once. It is a class of failure your prompts either invite or resist. A hallucination audit is the exercise of reading your prompt the way an adversary would, then reading your context the way a lazy reader would.
The four invitations
- Open-ended asks with no grounding: 'summarise everything important' with nothing to summarise against.
- Polite phrasing the model reads as licence: 'feel free to add detail' becomes fabricated detail.
- Missing refusal path: no instruction for what to do when the answer is not in the context.
- Confident closing demands: 'answer in full sentences' rewards filling gaps over admitting them.
Add the refusal, then the citation
Two additions cut most hallucination in retrieval-grounded prompts: an explicit instruction to say when the answer is not present, and a requirement to cite which passage supports each claim.
If the answer is not in the provided context, say "I don't have enough information to answer that."
For every claim, cite the source id in brackets, e.g. [doc_3]. If you cannot cite, do not make the claim.
Test with the gap input
Put a query whose answer is absent from the context into your eval set. The correct behaviour is a clean refusal. If the prompt invents an answer, the audit is not done — go back and tighten the grounding.
A model that never refuses is not a smart model. It is a confident one. Confidence without refusal is the whole problem.
