Mean reciprocal rank (MRR) measures how quickly a search system surfaces the first relevant result. It averages the reciprocal of that result’s rank across queries, so systems that consistently place relevant documents first or second score well. section 7.2.3
MRR focuses on the earliest relevant result rather than the full set of retrieved documents. Pair it with precision, which measures irrelevant results, and recall, which measures relevant documents the system fails to retrieve. section 7.2.3
Finding API integration case studies
If a user searches for “API integration case studies,” a high MRR indicates that a relevant case study appears near the top, allowing the user to find it without scrolling through irrelevant results. section 7.2.3
Use MRR to support practical search evaluation:
- Compare ranking changes and check whether optimization produces more immediately relevant results. section 7.2.3
- Evaluate retrieval separately from response generation when assessing a retrieval-augmented generation system. section 7.3.2
- Combine test results with click-through rate, query logs, dwell time, and bounce rate because production behavior can reveal ranking problems. section 7.2.3
