Relevance asks whether the answer addresses the question that was asked.
An answer can be perfectly true and still irrelevant.
Retrieved chunk relevance and final answer relevance are scored separately.
Low answer relevance often traces back to a poorly rewritten query.
It is usually scored by a rubric or by another model acting as judge.
Relevance asks whether the answer addresses the question that was asked.
It is easy to overlook because it sounds trivial.
An answer can be accurate, well grounded, well cited and beside the point.
Asked why a charge appeared, the model explains what the charge is.
That is a genuine failure and every other metric scores it well.
Chunk relevance and answer relevance are scored separately.
Chunk relevance asks whether retrieval returned anything useful.
Answer relevance asks whether the response used it to answer the question.
Low answer relevance often traces back to a bad query rewrite.
If the rewritten query drifted, everything after it answers the wrong question.
It is usually scored by a rubric or by a model acting as judge.
One approach has a model generate questions the answer would suit, then compares them with the real one.
Evasive answers are the common failure, since safety instructions can make a model hedge.
Watch relevance alongside faithfulness, because tightening one loosens the other.
A system that is always faithful and rarely relevant is a system nobody uses.
One question on this card. Every option explains itself once you pick, right or wrong.
18 concepts in this module, then a graded test with every answer explained.