# What is Reranking?

Source: https://mushbloom.uk/resources/glossary/what-is-reranking.html

By Mushbloom. Published and updated: 2026-10-12.

## Definition & overview

Reranking reorders an initial set of retrieved results using an additional relevance assessment. It can improve which passages are supplied to a model, but cannot recover evidence absent from the candidate set.

## Analogy & mental model

A researcher sorts a shortlist of documents before deciding which ones belong in the briefing.

Test retrieval and reranking separately, including questions where no suitable source exists.

## Why it matters

Initial retrieval may return many plausible passages. A second relevance stage can improve which evidence reaches the answer generator. It cannot repair a collection that lacks the needed source or a retriever that failed to include it.

## Practical business example

A support query returns several policy sections. Reranking prioritises the one covering the exact product and customer condition before generation.

Illustrative scenario, not a client case study or promised result.

## How it works

1. Retrieve a candidate set using keyword, vector or hybrid search.

2. Score candidates against the specific question with a suitable relevance model or method.

3. Supply the best permitted passages and preserve their source references.

## Core components

- Candidate set

- Relevance scoring

- Selected evidence

Better-looking relevance scores do not establish factual correctness or permission to use a passage.

## Reranking vs Retrieval

Retrieval creates the candidate set. Reranking reorders that set; evaluating the stages separately helps identify whether a missing answer was never retrieved or merely ranked poorly.

## Frequently asked questions

What does Reranking mean?
Reranking reorders an initial set of retrieved results using an additional relevance assessment. It can improve which passages are supplied to a model, but cannot recover evidence absent from the candidate set.

How is it different from Retrieval?
Retrieval creates the candidate set. Reranking reorders that set; evaluating the stages separately helps identify whether a missing answer was never retrieved or merely ranked poorly.

What is a sensible first step for a small business?
Test retrieval and reranking separately, including questions where no suitable source exists. Retrieve a candidate set using keyword, vector or hybrid search.

What limitations should I keep in mind?
Better-looking relevance scores do not establish factual correctness or permission to use a passage.

## Sources

Google Cloud: Retrieval-augmented generation: https://cloud.google.com/use-cases/retrieval-augmented-generation

No rankings, AI citations or commercial outcomes are guaranteed.