[KongchangAI]
Conceptany expected document in top-k / 任意命中

any-hit

A judgment rule in RAG retrieval evaluation in which a query counts as a hit if any one of the expected relevant documents appears in the Top-K results. In multi-document query scenarios it treats partial recall and complete recall as the same outcome, causing the metric to be inflated.

Timeline (last 90 days)

Oct 16

The default logic of most RAG evaluation tools is any-hit: as long as any document labeled as relevant appears in the Top-K results, it counts as a hit.

Unverified50%
Oct 16

When a query requires multiple pieces of evidence, any-hit records 'only getting part of the evidence' and 'getting the evidence in full' as the same result, leading to inflated metrics and distorted comparison between approaches.

Unverified50%
Oct 16

In deployment decisions, all-evidence should serve as the primary reference metric, because it is closer to the user's actual experience; any-hit can be retained as a diagnostic tool.

Unverified50%
Oct 16

If the any-hit score is high while the all-evidence score is low, the problem lies in coverage, and one may consider increasing Top-K, improving the chunk strategy, or introducing multi-hop retrieval; if both are low, it indicates a problem with the retriever's basic capability.

Unverified50%

All Facts (4)

Source Articles