Recall at K

What Is Recall at K?

Recall at K, also written Recall@K, measures the share of all relevant items found in the top K results. It evaluates ranked results from search engines, recommendation systems, code search, and retrieval-augmented generation, or RAG. A relevant item helps answer a query, and K is the number of top-ranked results checked. Recall@K does not measure irrelevant results or whether a useful item ranks first.

How to Calculate Recall at K

Calculate the score using a known set of relevant results for each query.

1. Define the documents, code files, products, or chunks relevant to a test query.
2. Retrieve and rank results, then keep the first K items.
3. Count relevant items in those results. These are true positives.
4. Divide that count by all known relevant items. If a RAG question has four supporting chunks and the top five contain three, Recall@5 is 3 divided by 4, or 0.75.

Relevant items not retrieved are false negatives, which lower recall.

Recall at K vs Precision at K vs Hit Rate

These metrics answer different questions about ranked retrieval.

Recall@K treats a relevant item at rank 1 the same as one at rank K, so it does not reward better ordering within the window.

Benefits of Recall at K for Search and RAG Evaluation

Recall is useful when missing source material could produce an incomplete answer.

• It identifies relevant documents or chunks a retriever misses.
• It helps teams tune embeddings, query rewriting, filters, and candidate limits.
• It supports evaluation for recommendations, semantic search, code search, and RAG.
• It makes retrieval testing concrete before evaluating generated answer quality.

Practical Limits of Recall at K

A strong score does not guarantee a good user result.

• It depends on complete, accurate relevance labels, which can be difficult for broad queries.
• Increasing K can raise recall simply by returning more items.
• It ignores ordering among relevant items within K.
• High recall can add noisy context, slow retrieval, and use excess model tokens.

Choosing K and Reading the Score

Choose K based on the results a user sees or the chunks a downstream model can use well. Report several values, such as Recall@3, Recall@5, and Recall@10. Segment results by query type, and pair recall with precision and answer-quality checks. This approach helps when planning app design and development for AI features.

Recall at K in Emergent Apps

Teams building AI-powered search, knowledge bases, recommendation features, or RAG workflows in Emergent can define relevant test results before measuring Recall@K. This creates a repeatable way to validate retrieval as an app evolves. Explore related concepts in the Emergent glossary, then use recall with user testing to improve AI retrieval.