Polp
← Back to glossary
RAG and search

Reranking

A second-stage retrieval process that takes a small candidate set and reorders it using a more precise, query-aware relevance model.

Business example

A first search returns twenty HR passages mentioning leave. A reranker moves the approved parental-leave clause above general meeting notes before the answer model receives context.

How it works

  1. 1.Retrieve a broad candidate set quickly with lexical, vector, or hybrid search.
  2. 2.Score each candidate again against the exact user query with a more precise model.
  3. 3.Reorder the candidates and keep only the strongest passages.
  4. 4.Send the selected passages and their source metadata to the generation stage.

Common misconceptions

Reranking searches the entire document collection from scratch.
It normally works on a limited candidate set returned by a faster first-stage retriever.
A higher reranker score proves that a passage is factually correct.
The score estimates relevance to the query; authority, freshness, permissions, and factual support still need separate controls.

Related concepts

Retrieval-Augmented Generation (RAG)Semantic searchEmbeddingsVector database

Related reading

The new AI SaaS metric: answer quality, not document count
RAG is not uploading PDFs: permissions, freshness, traceability, and reliable sources

Sources

Reviewed:

Reviewed by: Javier Chulvi Bernad · LLM Engineer · Madrid