Reranking: meaning, examples and practical use
Reranking evaluates and reorders an already selected set of results against a query. It can apply a more detailed relevance assessment after initial retrieval. Reordering candidates is not the same as searching the whole collection again or proving the sources true.
Also known as: reranker · re-ranking
At a glance
- Reranking operates on candidates already retrieved.
- A relevance score is not factual verification.
- A missing necessary document must be found by another retrieval step.
On this page

Why a second ordering?
Initial retrieval can quickly select plausible candidates from a large collection. A later stage can examine the relationship between the query and each shortlisted passage more closely. Google’s ranking documentation describes this reassessment of already retrieved records. Ranking documentation
The objective is a more useful order for the task, rather than arbitrarily reshuffling the same list.
Inputs and outputs
Inputs can include the question and several document titles or passages. Outputs can include a new order and relevance scores. Actual fields, limits and score ranges depend on the implementation.
Reranking is not the whole RAG process. It helps select information; writing the answer is another operation. This internal relevance order also differs from a company’s position in Google Search.
When comparing systems, ask which candidate set each received. A better outcome from a different input list cannot automatically be attributed to the reranker alone.
Example: a venue’s booking conditions
Imagine a venue knowledge base receiving “How late can I change the guest count?” Initial retrieval returns its introduction, quote page and booking conditions. A more detailed assessment may promote the conditions containing the actual deadline.
If that document never entered the shortlist, reranking cannot select it. Investigate the earlier search or the document’s availability instead.
What to evaluate
Compare original and revised orders using questions with known necessary sources. Include outdated information, misleading near-matches and similar names. More detailed scoring may require additional time or computation, so measure whether it helps.
Answer groundedness remains separate. A response faithfully using an old rule can still be outdated. Ranking cannot replace checking source freshness.