The fraction of retrieved items within a result set that are judged relevant to the user's query or information need, calculated under specified relevance criteria and retrieval cutoff.

Principle

Precision measures the accuracy of the retrieved set: raising precision typically requires stricter filtering or better ranking, which can reduce recall unless retrieval relevance improves overall.

Demonstration

Illustrative scenario: a result set of 50 returned items includes 40 judged relevant under the chosen cutoff → precision = 0.80; recognition: 80% of returned items meet the relevance criterion; action: developers adjust ranking to raise the proportion of relevant items in top results.

Misapplication

Treating high precision as equivalent to comprehensiveness; the error is conflating accuracy of returned items with coverage of all relevant items in the collection.

Consequence

High precision reduces the user's effort to find relevant items within returned results and is valuable in contexts where users inspect few items; however, over-optimizing precision can omit relevant material and harm recall-dependent tasks.

Reversal

When the task requires exhaustive retrieval (e.g., legal discovery, systematic review), prioritizing precision over recall can be counterproductive; conversely, in narrow-answer tasks precision is paramount.

Boundary

Clearly within: proportion of relevant items among the top-N retrieved results under a defined cutoff. Boundary case: graded relevance where partial relevance complicates binary precision computation. Clearly outside: measures focusing on rank-discounted utility across all positions (e.g., MAP or NDCG) rather than simple fraction of relevant items.

Semantic Tension

Precision (accuracy of retrieved items) ↔ Recall (coverage of relevant items); improvements in one often trade off with the other absent holistic improvements in relevance modeling.

Synthesis

Precision quantifies how much of a retrieved set is relevant and directly relates to user effort reviewing results; it must be balanced with recall and ranking considerations according to task requirements.