On this page
The core idea
Precision describes how much of the retrieved set is relevant; recall describes how much of the relevant material was found. In practice, the definition of relevance and the available labels matter. Use these measures as diagnostic tools alongside question-level inspection, because a single missing passage may contain the condition needed for a correct answer.
A practical approach
Step 1
Define relevance for your task and label supporting passages for a set of questions. Record when the full relevant set is not known.
Step 2
Compare retrieval configurations at a consistent result cutoff. Inspect which useful passages were missed and which irrelevant ones were added.
Step 3
Evaluate the downstream answer too. Better retrieval metrics do not automatically guarantee a supported and useful response.
Worked example
If a query returns five passages and three are relevant, precision at that cutoff is three fifths. If the labelled collection contains four relevant passages and three were found, recall against those labels is three quarters. These toy numbers illustrate the definitions, not a target for every application.
A mistake to avoid
Incomplete relevance labels can make a useful result look like a false positive. Review unexpected results instead of trusting the metric mechanically.
Try it yourself
Label a small document set for five questions and calculate both measures by hand. Explain one case where the aggregate score hides an important missing condition.
Your implementation checklist
- Relevance is defined explicitly.
- Comparisons use a consistent cutoff.
- Label completeness is understood.
- Question-level failures remain visible.
Share or cite this guide
Innovate with Ajay Editorial (2026-10-04). Understand Retrieval Precision and Recall with Examples. https://innovatewithajay.com/blogs/measure-retrieval-precision-and-recall-with-examples
Editorial approach, sources and corrections
Apply this guide to your own work: build a project brief with inputs, test cases and review boundaries.