← All questions
EvaluationMedium· RAG· Grounding· Metrics

Our citation coverage is 100% and users still tell us t...

Citation coverage measures whether the relevant document was retrieved, but it does not measure whether the model actually used that document to generate a factually supported sentence. The fix is to introduce an entailment verification step. This involves checking if the generated output is logically supported by the cited span. This check can be computationally expensive, so it should be implemented as an offline sample-and-rerank process to manage costs.

Code playground (Python)