Mean reciprocal rank (MRR) evaluates how quickly a ranked retrieval system surfaces the first correct or relevant result. For each query, score 1 divided by the first relevant result's rank, then average those scores across queries. section 3.2.1 section 14.1.1
MRR rewards early retrieval sharply: rank 1 contributes 1.0, rank 2 contributes 0.5, and rank 50 contributes 0.02. This makes it useful when downstream generation sees only a few ranked documents or attends unevenly across the context. section 3.2.1 section 14.1.1
Two-query calculation
If the first relevant compliance document appears at rank 3 for one query and rank 1 for another, MRR = (1/3 + 1/1) / 2 = 0.667. An MRR of 1.0 means the top result is always relevant. section 14.1.1
Use MRR alongside other retrieval metrics because it emphasizes the first hit rather than the whole ranked set.
- Recall@k records whether any correct document appears within the first k positions, even when it is buried; MRR penalizes that delay. section 3.2.1
- NDCG@k considers multiple relevant documents and discounts their positions, whereas MRR considers only the first relevant result. section 14.1.1
- Interpret gains with a baseline and check accuracy, latency, index size, dataset splits, and statistical evidence rather than treating MRR alone as sufficient. section 3.2.1
The equation expresses MRR as the average reciprocal rank of the first relevant result across the query set. section 14.1.1




