Evaluation
Hallucination
A hallucination is model output that is fluent, confident and not supported by the source material or by fact.
Fabricated numbers are the highest-signal case: a figure in the answer that never appeared in the provided context is almost always invented, and it is far more damaging than a vague sentence because people act on it.
Retrieval reduces hallucination by supplying grounding material, but does not eliminate it — a model shown four passages can still produce a fifth claim.
In practice
The reliable trigger is a question the context cannot answer combined with no permission to say so. A model given no useful retrieved material and no defined fallback will produce something confident and well-formed. Deciding in the prompt what happens when the answer is not present removes most of it, and costs one sentence.
Common questions
How do I stop a model from making things up?
Supply the facts through retrieval, and state explicitly what to do when the answer is not in the context. A model with no useful material and no permission to say so will produce something confident.
Do citations prove an answer is grounded?
No. Models produce citation-shaped text for sources they did not use. Verify that each cited chunk is actually in the retrieved set — cheap to implement, and the difference between a citation and a decoration.