Recall@k

Recall@k asks: of everything known to be relevant, what fraction appeared in the first k results?¹

Recall@k = relevant results in the first k / all known relevant results

Tiny example

Two passages are labeled relevant. The top five results contain one of them.

Recall@5 = 1 relevant result found / 2 known relevant results = 0.50

High recall gives the model a better chance of seeing all the necessary evidence. It does not guarantee that the top results are clean or well ordered.

FDE note

Recall needs a trustworthy set of relevance labels. A single “gold passage” per question can make the denominator misleading. Read it beside Precision@k, not in isolation.

Questions with no relevant passage have a zero denominator, so evaluate them in an unanswerable-query or abstention suite instead of silently assigning a Recall@k score.

Footnotes

1. Google for Developers, Recall@k. Microsoft describes the denominator as all possible relevant items in its RAG information-retrieval guidance. ↩