Definition
Mean Reciprocal Rank (MRR) is a retrieval evaluation metric that calculates the average of the reciprocal ranks of the first relevant document across a set of queries. In RAG pipelines, it measures the retriever's effectiveness at placing the most pertinent ground-truth context at the very top of the result list, where a rank of 1 yields a score of 1.0 and a rank of 10 yields 0.1.
A search relevance metric for RAG evaluation, not a SaaS business metric for revenue.
"A target where only your best-placed arrow counts; a bullseye is a perfect score, but the value drops sharply the further that single best arrow is from the center."
- NDCG (Normalized Discounted Cumulative Gain)(Alternative Metric for multi-document relevance)
- Recall@K(Complementary Metric)
- Top-K Retrieval(Prerequisite)
Conceptual Overview
Mean Reciprocal Rank (MRR) is a retrieval evaluation metric that calculates the average of the reciprocal ranks of the first relevant document across a set of queries. In RAG pipelines, it measures the retriever's effectiveness at placing the most pertinent ground-truth context at the very top of the result list, where a rank of 1 yields a score of 1.0 and a rank of 10 yields 0.1.
Disambiguation
A search relevance metric for RAG evaluation, not a SaaS business metric for revenue.
Visual Analog
A target where only your best-placed arrow counts; a bullseye is a perfect score, but the value drops sharply the further that single best arrow is from the center.