Precision

What is precision?

Back to the library. This time the system returns 100 books for "machine learning" — but 40 of them are novels and cookbooks that have nothing to do with it. Precision measures accuracy: of all the results returned, what fraction is actually relevant.

How do precision and recall work together?

A built-in tension
Want to miss nothing? Loosen the search and junk floods in. Want zero mistakes? Tighten it and you start dropping good results. Precision and recall sit on a seesaw — push one up, the other dips.
Pick a side by context
In search and recommendations, users only read the first few results, so precision wins. In medical screening, one miss is a disaster, so recall wins. Neither is inherently better — it depends what you care about.

How do you calculate precision?

The numerator is the number of returned results that are truly relevant; the denominator is everything the system returned. Return 100 items, 60 of them relevant, and precision is 60%. It answers a simple question: how much of what I gave you is actually good.

How do you raise precision?

Match more strictly
Exact keywords and filtering out low-relevance candidates cut the filler.
Add scoring and a threshold
Score candidates and keep only those above the line to push accuracy up.
Use a reranker
Roughly retrieve with a cheap method, then re-rank with a stronger model — the classic way to boost precision.

Bottom line: precision asks "is what I found actually good?" — it decides how much of your answer is truly useful.