Skip to content
ML Academy

Introduced on day 51

Precision@k

The share of the top k results that are relevant; its maximum is below 1 whenever a query has fewer than k relevant passages.

In detail

Precision@k = (relevant passages in the top k) / k. With one or two relevant passages and k = 5 the best possible value is 0.2 to 0.4, so a low precision@k is often a property of the labels, not of the retriever. It is most useful for comparing runs at the same k, or for deciding how much irrelevant text the generator receives.