Sort Passages takes a short list of search snippets or retrieved paragraphs and puts the ones that actually answer your question at the top.
Use it after a search or a first-pass retrieval, not over a whole library. Each candidate gets a relevance score. You then keep the top few and send only those to the answering model.
This is cheaper than pasting every snippet into chat. A typical call finishes in about a tenth of a second.
What you can do
- rerank retrieved paragraphs before the answering model sees them
- rerank web-search snippets against the original question
- Drop clearly irrelevant candidates instead of reading them
Who it's for
Search UIs, retrieval pipelines, and agents that found more passages than they should paste.
How to use it
- Fetch a short list in code (max 40)
- Call rerank with the question and candidates
- Send only the top few passages to the answering model
Getting started
Pass a refund-window question and three snippets. The policy passage should beat shipping and product copy.
