Precision is a search-quality metric that measures the fraction of returned results that are genuinely relevant. If a query returns ten results and seven of them are truly relevant, the precision is seventy percent. It answers the question: of what the system gave me, how much was actually useful?
Precision is often discussed alongside recall, its counterpart. Recall measures how many of all the relevant items the system managed to find, while precision measures how clean the returned set is. The two trade off against each other — returning more results tends to raise recall but lower precision, since the extra results are more likely to be marginal. The right balance depends on the application.
In retrieval-augmented generation, precision is especially important because irrelevant retrieved chunks waste the model’s limited context and can mislead its answer. This is why a re-ranking step is often added: a fast first stage retrieves a broad set with good recall, and a more accurate re-ranker then improves precision by promoting the genuinely relevant items to the top before they are used.