Skip to content
ML Academy

Introduced on day 51

Recall@k

The share of a query's labelled relevant passages that appear in the top k results; for RAG, k is the number of passages passed to the model.

In detail

For one query, recall@k = (relevant passages in the top k) / (all labelled relevant passages). A looser "hit rate" only asks whether any relevant passage appears; the two coincide when each query has one relevant passage. Recall ignores order inside the top k, so a passage at rank 1 and at rank k count the same. A recall@100 figure says nothing about a prompt that receives five passages.