Mean Reciprocal Rank (MRR)
Top-1 accuracy is all-or-nothing: a ranker that puts the right answer second scores exactly zero. MRR grades the whole ranking instead; take the reciprocal of the position where the correct answer landed (1st → 1.0, 2nd → 0.5, 4th → 0.25, missing → 0) and average it across all queries. Handy intuition: 1 ÷ MRR ≈ where the right answer typically sits in the list. I met it while comparing a lexical name-matcher against a sentence-transformer for ranking schema-field suggestions: their top-1 counts were close, but MRR (0.48 vs 0.60) revealed the model was missing higher up the list, which is the metric that actually matters when a human confirms answers from a ranked list