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.
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.
- 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
ragsemantic-searchchunking- etymology_origin
- — verify-against-primary-at-build ↗ https://en.wikipedia.org/wiki/Learning_to_rank — reranking builds on the learning-to-rank tradition in information retrieval
- related_to
ragsemantic-searchchunkingembeddings- 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
last verified · by Özden Erdinc