Recall

Of all the documents that should have been returned, how many did the system actually find? That's recall. If your corpus has ten documents relevant to a query and the system retrieves seven of them, recall is 0.7. The other three are invisible to the user--they exist, they're relevant, and the system missed them.

Like precision, recall is reported at a cutoff: Recall@10 means "of all relevant documents in the corpus, how many showed up in the top 10 results?" A system with high Recall@10 is finding most of what matters without making the user dig. A system with low recall is hiding relevant content, and no one knows it's missing because you can't miss what you never saw.

High recall and high precision pull in opposite directions. Casting a wider net (returning more results) catches more relevant documents but also drags in more noise. Tightening the filter cuts noise but risks dropping relevant content. Every retrieval system lives somewhere on this tradeoff, and the right balance depends on what the system is for. A medical information system should favor recall (don't miss the relevant study). A customer FAQ chatbot should favor precision (don't confuse the user with tangentially related answers).

Why it matters for writers: Low recall often means your documents are failing to match queries they should. The content is relevant but the system can't tell. Common causes: content that uses different terminology than the queries, documents too long to embed meaningfully, or missing metadata that would help the retrieval system understand what the document is about as opposed to what it says.

Related terms: Precision · Semantic Search · Chunking