---
language: en
title: "Reranking"
description: "Reranking reorders results retrieved for a query. Learn how to compare the lists, assess relevance, and recognize the limits of this search stage."
canonical: "https://tropk.ai/en/resources/glossary/reranking"
last_modified: "2026-10-12T16:13:07.470Z"
---

# Reranking

Reviewed on 2026-10-06.

## What it means

Reranking is the reordering of results that have already been retrieved, using an additional relevance assessment for the query. The initial search efficiently gathers candidates; a second stage may use a more costly model to score a smaller set. The item being reordered may be a document or a passage, depending on the application design and its [retrieval](/en/resources/glossary/information-retrieval) unit.

## Why it matters for marketing

For marketing, this distinction makes it possible to investigate answers supported by unsuitable pages without attributing every failure to the published content. A useful source may be among the candidates but receive low priority. When evaluating a search solution, the team should ask for evidence of the order before and after reranking, along with the additional time needed to obtain the result.

## Example in practice

In a hypothetical test, a course platform retrieves forty passages for a question about cancellation after classes begin. The specific clause is in position eighteen. After reranking, it moves to second place, ahead of the commercial overview. The team compares the lists and decides to test generation with the clause included in the context.

## How to apply or verify

Prepare queries with defined relevance criteria and keep the candidates the same when comparing configurations. Record the initial order, final order, scores, and latency. Pay particular attention to relevant passages that dropped in position. The technical lead should measure the improvement on the evaluated set and check whether the cost allows the system to meet real-world usage requirements.

## Important distinctions

Reranking is not the same as running the search again or writing the answer. It also does not have to use a [cross-encoder](/en/resources/glossary/cross-encoder), although that is a common method in retrieval workflows. A better-ordered list does not prove that the generated answer is faithful. Scores produced by the reranker belong to the model used and should not be automatically compared with those from the initial search.

## Sources and references

- [Retrieve & Re-Rank — Sentence Transformers documentation](<https://www.sbert.net/examples/sentence_transformer/applications/retrieve_rerank/README.html>) — Sentence Transformers

## Continue a leitura

- [Semantic search](</en/resources/glossary/semantic-search>)
- [Candidate set](</en/resources/glossary/candidate-set>)
- [Cross-encoder](</en/resources/glossary/cross-encoder>)

## About the author

[Vinicius Pacheco](</en/team/vinicius-pacheco>)

Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content.

---

[All pages in Markdown](<https://tropk.ai/sitemap.md>)