Recall

What is recall?

Picture yourself searching a huge library for books on machine learning. The system hands you 10, but 40 more relevant books are still sitting on the shelves. Recall measures whether it found everything — of all the documents that are actually relevant, what fraction did the system manage to return.

How do you calculate it?

What's in the numerator and denominator
The denominator is the total number of relevant documents; the numerator is how many of those the system actually retrieved. Divide one by the other and you get recall. If there are 100 relevant articles and the system finds 80, recall is 80%.
It pairs with precision
Recall asks "did I miss anything?", precision asks "was I wrong about any of these?". One cares about completeness, the other about correctness — and they usually trade off against each other.

Why recall matters

In medicine, law and risk screening, missing something costs far more than returning a few extras. A missed diagnosis or a missed piece of evidence can be catastrophic. So in those fields, people would rather get too many results than too few — they push recall up on purpose.

How do you raise recall?

Loosen the matching
Widen the search, allow synonyms and fuzzy matches, and you'll pull in more relevant content.
Multi-way retrieval
Run several strategies at once — keywords, vectors, semantics — then merge, and you cut the misses significantly.
Hybrid search
Fuse classic keyword results with vector results to get both breadth and accuracy.

Bottom line: recall asks "did I find everything I should have?" — it decides whether you miss the signal that matters.