Recall@k
Recall@k is a search metric that measures the share of all relevant results that appear in the top k results. If a query has 8 relevant pages and 6 of them are in the top 10, recall@10 is 0.75. When each query has a single target, as in known-item search, recall@k is the share of queries whose target appears anywhere in the top k.
Recall is one of the two original measures of information retrieval, alongside precision, and was used in the Cranfield experiments of the 1960s. Recall@k matters for AI agents because an agent usually reads only the first few results, and a relevant page at position 40 is unlikely to be read.
Example
A benchmark has 100 queries, each with one target paper. The target appears in the top 10 for 72 queries, so recall@10 is 0.72. If 51 of those targets are in the top 5, recall@5 is 0.51.
How it works in NEEDLE
NEEDLE reports recall@10 for Scholar and recall@5 for Legal, matching paper identifiers (arXiv ID, DOI, PMID) or legal citations and docket numbers, and answer-recall@5 for Finance, where a hit means the gold fact appears in a top-5 title or snippet. See the live NEEDLE results.
FAQ
What is the difference between recall@k and precision@k?
Recall@k is the share of all relevant results that appear in the top k, and precision@k is the share of the top k that are relevant. A list can have high precision@5 and still low recall@5 when many relevant pages exist.
Which k should you use for recall@k?
Choose k to match how many results the user or agent actually reads. For an agent that reads the top five snippets, recall@5 says more than recall@100.
Related terms
Written by Keenable.
Published
