Also known as: P@K, precision at K
Precision@K is the fraction of the top-K returned documents that are relevant. The classical IR metric retrieval moved away from in favor of NDCG, but still the right choice when every position in the result list carries equal weight.
Precision@K is the fraction of the top-K results that are relevant:
For a single query, if 7 of the top 10 results are relevant, Precision@10 = 0.7. Average across queries to get a corpus-level number.
It’s the natural “of what I returned, how much was good?” metric — the mirror image of Recall@K , which asks “of what was good, how much did I find?”.
Classical IR papers from the 1990s lean heavily on Precision@K and MAP . Modern retrieval evaluations lean on NDCG@K instead. The reason is position weighting.
Precision@10 counts a relevant doc at position 1 and a relevant doc at position 10 equally. Users don’t. Eye-tracking and click data show attention drops sharply with rank. NDCG bakes that into the metric with a logarithmic discount; precision treats the result list as an unordered set of K items.
For ranked-list consumers — users scrolling search results, RAG pipelines passing top-K to an LLM that weights them by order — NDCG is the more honest metric. Precision is only correct when every returned position is treated equally downstream.
Outside those cases, NDCG@K or MRR is usually the better choice.
Order matters in the prompt’s structure but rarely in the LLM’s attention across short contexts. With 5 chunks of ~500 tokens each, the generator reads all 2500 tokens; the position-1 chunk doesn’t get a
The two measure orthogonal things:
A trivial system returning every document has perfect recall and terrible precision. A system returning only its single most confident result has high precision and terrible recall. Tuning K (or a score threshold) sweeps the curve between them; F1 score is the harmonic mean if you want a single number.
Pick the metric that matches how your downstream consumer reads the result list. Precision when they read everything; NDCG when they read top-down.
Precision@K treats position 1 and position K identically — a relevant doc anywhere in the top-K counts the same. Real users read top-down and tire fast, so position 1 is worth much more than position 10. NDCG bakes that with logarithmic discount; Precision doesn't. For ranked-list consumers, NDCG is the more honest metric.
When every position in the top-K is consumed equally — a classifier emitting K candidates that all get manual review, a retrieval audit where humans triage every result, or a multi-document RAG step where all K passages are concatenated as context. There the user reads them all, so position weighting is misleading.
Precision rises as K shrinks (you only keep the most-confident candidates) and falls as K grows. Recall does the opposite. The classical curve sweeps K (or a score threshold) and reports both — area under it is closely related to MAP. Tune K based on the downstream consumer's tolerance for noise.
