Mean reciprocal rank (MRR) measures how quickly a retrieval system surfaces the first relevant result. For each query, it uses the reciprocal of that result’s rank, then averages across queries. section 6.2.3
A relevant result at rank 3
If the first relevant document appears at rank 3, its reciprocal rank is 1/3, approximately 0.33. This makes MRR useful when one good document is often sufficient to answer a question. section 6.2.3
Choose MRR according to how retrieved context is used:
- Use MRR when any one relevant document can answer the query. section 6.2.3
- Prefer Precision@k when a fixed number of documents enters the context window. section 6.2.3
- Use nDCG when all relevant documents matter and relevance has graded levels. section 6.2.3
- Interpret MRR alongside precision and recall rather than treating it as a complete retrieval diagnosis. section 6.5




