What is recall@k? Retrieval metrics for RAG explained

Also called recall at k, R@k, retrieval recall

Definition

Recall@k is a retrieval metric that measures the fraction of a query's relevant documents that appear in the top k results, averaged over a set of test queries.

Recall@k is a retrieval metric that measures how many of the relevant documents for a query appear in the first k results. It is the relevant documents found in the top k divided by all relevant documents for that query, averaged across a test set. In a RAG pipeline, recall@k caps answer quality, because a passage that never reaches the top k never reaches the model. It is the first number to compare when choosing between chunking strategies, embedding models, or retrieval settings on the same labeled queries.

How to compute recall@k#

Recall@k is computed per query and then averaged. For each test query, count the labeled relevant documents that appear in the top k results and divide by the total number of relevant documents for that query. With one relevant passage per query, recall@k reduces to the hit rate, the share of queries with at least one hit.

For example, a query has four relevant chunks, and the retriever's top 10 contains three of them, so recall@10 for that query is 3 / 4 = 0.75. A second query with two relevant chunks, both in the top 10, scores 1.0. Recall@10 for the two-query set is 0.875.

Recall rises as k grows, so compare systems at the same k and the same token budget. A first-stage retriever that feeds a reranker is judged at a large k, such as 50 or 100, because the reranker can only reorder what it receives.

A query with more relevant documents than k can never reach 1.0. The BEIR benchmark (Thakur et al., 2021) calls plain recall counterintuitive in this case and caps it.

Anthropic's contextual retrieval study reported its results as a failure rate, defined as 1 minus recall@20.

Recall@k vs precision@k, MRR, and nDCG#

Recall@k asks whether the relevant documents made the list. Precision@k asks how much of the list is relevant. MRR asks how high the first relevant result ranks. nDCG@k scores the whole ordering, giving more credit to relevant results near the top. Track recall@k alongside at least one rank-aware metric.

The Elasticsearch ranking evaluation API implements all four and notes that precision and recall at k ignore where in the top k a relevant result sits.

Metric What it measures Rank-aware Typical use in RAG
Recall@k Share of relevant documents found in the top k No First-stage retrieval, chunking comparisons
Precision@k Share of the top k that is relevant No How much context budget goes to noise
MRR Average of 1 / rank of the first relevant result Yes Questions with one right passage
nDCG@k Graded relevance discounted by position, normalized to the ideal order Yes Final ranking after reranking

BEIR reports nDCG@10 as its main metric because it handles both binary and graded relevance judgments, and the late chunking paper uses the same measure.

Retrieval recall@k differs from the recall of an approximate nearest neighbor index, which compares the index with exact vector search and needs no relevance labels. Pair retrieval metrics with a generation metric such as faithfulness to see whether a bad answer started in retrieval or in generation. The chunking strategies guide includes a short Python evaluator for recall@k and MRR.

Cite this definition

The AI Engineering Editors. "What is recall@k? Retrieval metrics for RAG explained." The AI Engineering, updated Nov 06, 2026. https://theaiengineering.com/glossary/recall-at-k

Title
What is recall@k? Retrieval metrics for RAG explained
Author
The AI Engineering Editors
Publisher
The AI Engineering
Updated
View as Markdown

Part of the RAG and vector search hub · Updated