The fraction of all items judged relevant to a particular query or information need within a defined collection that are successfully retrieved by the search system under a specified relevance judgment and retrieval cutoff.

Principle

Recall quantifies coverage of relevant material: increasing recall requires retrieving a larger share of relevant items, often at the cost of lower precision unless retrieval quality improves.

Demonstration

Illustrative scenario: within a closed collection where 100 items are judged relevant to a topic, a retrieval run returns 80 of those items within the result cutoff → recall = 0.80; recognition: the system covers 80% of known relevant items under the defined conditions; action: developers tune indexing or query expansion to increase recalled items.

Misapplication

Assuming high recall alone implies good overall retrieval performance or user satisfaction; the error is ignoring precision, ranking, and the user's tolerance for non-relevant items.

Consequence

A system with higher recall reduces the risk of missing relevant documents and is preferable in tasks where omissions are costly (e.g., systematic review), but may increase the volume of non-relevant items requiring inspection.

Reversal

Recall is not computable or meaningful when the set of relevant items is undefined (open-web searches, evolving relevance criteria) or when relevance is graded rather than binary without an agreed threshold.

Boundary

Clearly within: evaluation in a closed test collection with an externally defined set of relevant documents. Boundary case: interactive search where relevance depends on evolving user intent—recall estimates become approximate. Clearly outside: measures of ranked-list quality that account for graded relevance and position (e.g., NDCG) rather than simple recall.

Semantic Tension

Recall (coverage of relevant items) ↔ Precision (accuracy of retrieved items); retrieving more relevant items tends to reduce precision absent other improvements.

Synthesis

Recall measures the proportion of relevant items a system retrieves under specified conditions; it is essential for coverage-sensitive tasks but must be interpreted alongside precision, ranking, and the feasibility of reviewing extra results.