Precision at K
Also called Precision@K
Of the top K results you returned, how many were actually relevant.
Think of it like
Of the ten job candidates you shortlisted, how many were genuinely qualified.
Example
If 8 of your top 10 retrieved chunks are on-topic, precision@10 is 0.8; the other two are noise the model must ignore.
How it actually works
Precision@K measures how clean your top results are. In RAG it matters because irrelevant chunks waste context tokens and can distract the model into wrong answers. Precision and recall pull against each other, which is why pipelines retrieve broadly for recall, then rerank to push precision up in the final few passed to the model.
For product teams
Low precision means the model wades through junk; it shows up as vague or off-topic answers.
For engineers
|relevant ∩ top-K| / K; typically maximized via a reranking stage after high-recall retrieval.
Related
- Recall at K — The recall counterpart.
- Reranking — Reranking exists to raise it.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome