Hit Rate@K

Definition

Hit Rate@K (HR@K) measures whether at least one relevant document appears in the top K results for a given query. It is a binary, per-query metric.

Hit(query) = 1  if any relevant doc is in top K results
           = 0  otherwise

HR@K = mean(Hit) across all queries

Unlike NDCG or MRR (which reward ranking relevant docs higher), Hit Rate@K only asks: “did we find anything useful at all?”

When to Use

ScenarioWhy HR@K fits
RAG / question answeringRetriever must surface at least one answer passage
Recommendation systemsAt least one good recommendation above the fold
AutocompleteAt least one correct suggestion in top K
Retrieval health checkFast sanity check — HR@10 < 0.7 = retrieval is broken

HR@K is especially common in RAG evaluation: if the retriever misses all K candidates, the LLM has no chance of answering correctly.

Hit Rate vs. Other Metrics

MetricQuestionGranularity
HR@KWas anything relevant found?Binary per query
P@KWhat fraction of top K are relevant?Continuous per query
MRRHow far down is the first relevant doc?Position-aware
NDCG@KHow well are relevant docs ranked overall?Graded + position-aware

HR@K is the weakest signal (binary, no position info) but the fastest diagnostic.

Typical Thresholds

HR@KInterpretation
> 0.90Healthy retrieval; ranking is the bottleneck
0.70–0.90Retrieval gaps on tail queries
< 0.70Retrieval is broken or index coverage poor

HR@K vs. MRR

Both are per-query binary in spirit, but differ in reward:

  • HR@K: cares only that something relevant landed in top K
  • MRR: rewards finding that relevant doc as early as possible

A system with high HR@10 but low MRR is retrieving relevant docs but burying them at ranks 8–10.

Articles