MRR
also: mean reciprocal rank
Mean Reciprocal Rank: rewards how high the FIRST correct result appears, averaged over queries.
MRR (Mean Reciprocal Rank) scores a ranker by the position of the first relevant result: for each query the score is 1 divided by the rank of that first hit (1 if it is first, 1/2 if second, 1/3 if third), averaged over all queries.
Worked example: if the right answer appears at position 1 for one query and position 4 for another, MRR = (1/1 + 1/4) / 2 = 0.625; getting the answer to rank 1 matters far more than moving it from 5 to 4. Gotcha: MRR only cares about the FIRST relevant result, so it is ideal for known-item search (one right answer, like a document lookup) but a poor fit when many results are relevant — use NDCG or recall@k there.