Detecting and reducing hallucination

Faithfulness is measurable

Split an answer into claims and check each against the retrieved passages. Automating that check with a grader model gives you a faithfulness score you can track across releases.

Most hallucination in a RAG system is a retrieval failure in disguise: the passage was not there, and the model filled the gap. Check recall before rewriting the prompt.

Design for refusal. If saying "the documents do not cover this" is treated as a failure, the system will learn to guess instead, which is strictly worse.