Recall@k

Recall@k grades a retrieval step. For each question, the system returns a ranked list of candidate passages; Recall@k asks whether the correct one appears in the top k, and the score is the share of questions where it did. Recall@5 uses the top 5, Recall@10 the top 10.

It matters because a language model only reads the handful of passages you hand it. If the right one sits below the cutoff, the model never sees it and cannot answer, however capable it is. A high Recall@k means retrieval put the answer where the model will actually use it. It grades the fetching, not the final answer; groundedness covers that second half.