# Reranking

> A second retrieval stage that re-scores an initial set of candidate passages with a more precise model, promoting the most relevant to the top.

_The Agentic Web Lexicon · /glossary/reranking · [JSON](/api/glossary/reranking) · [all The Agentic Web Lexicon](/glossary)_

- **term:** Reranking
- **category:** knowledge-memory
- **short_def:** A second retrieval stage that re-scores an initial set of candidate passages with a more precise model, promoting the most relevant to the top.
- **long_def:** Reranking improves retrieval precision: a fast first stage (vector or keyword search) returns many candidates, then a slower, more accurate cross-encoder re-scores each against the query and reorders them. The agent then reads only the top few, so quality rises without a larger context budget.
- **see_also:** rag, semantic-search, chunking
- **etymology_origin:** — (verify-against-primary-at-build)
- **related_to:** rag, semantic-search, chunking, embeddings
- **contrast_with:** Unlike first-stage vector search optimized for recall and speed, a reranker optimizes for precision — it is slower but far better at ordering the final few results.
- **example:** A search returns 50 candidate passages; a reranker re-scores them and moves the 3 truly relevant ones to the top, which is all the agent reads.
- **source:** https://en.wikipedia.org/wiki/Learning_to_rank
- **status:** active
- **why_it_matters:** Reranking is how an agent spends a small context budget on the highest-signal passages — the retrieval-side counterpart to writing content an engine can rank cleanly.
- **sameAs:** https://en.wikipedia.org/wiki/Learning_to_rank
- **bridge_entity:** rag
- **last_verified:** 2026-07-06
- **md_twin:** /glossary/reranking.md