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 Reranking 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 Reranking works
- Retrieve a candidate set using keyword, vector or hybrid search.
- Score candidates against the specific question with a suitable relevance model or method.
- Supply the best permitted passages and preserve their source references.
Core components & mechanisms
- 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 & further reading
Google Cloud: Retrieval-augmented generationOriginal Mushbloom explanation and illustrative examples. References support the underlying topic, not every business recommendation. Check current documentation before implementation.
Turn the concept into a practical decision
Discuss your website or business workflow before choosing tools or implementation. Start with the problem, approved information and appropriate human oversight.
Book a free 20-minute audit →Explore the relevant serviceNo rankings, AI citations or commercial outcomes are guaranteed.