---
language: en
title: "Cross-encoder"
description: "A cross-encoder evaluates a query and passage together to score relevance. Learn how it is used in reranking, what to test, and its time and input limits."
canonical: "https://tropk.ai/en/resources/glossary/cross-encoder"
last_modified: "2026-10-12T16:13:05.539Z"
---

# Cross-encoder

Reviewed on 2026-10-06.

## What it means

A cross-encoder is a model that receives a query and a document or passage together, allowing it to evaluate their relationship during processing. In search tasks, it can assign a relevance score to each pair. This joint evaluation is often used to rerank candidates, since repeating the process across a large collection can be too costly.

## Why it matters for marketing

For marketing, this matters when the difference between two similar passages depends on a specific condition in the question. Joint evaluation can help select the relevant passage, but this should be demonstrated with business examples. When requesting this step, include relevance criteria and response-time limits rather than accepting only a promise of greater accuracy.

## Practical example

In a hypothetical test, accounting software is asked whether a digital certificate is needed to export reports. One passage requires a certificate to submit obligations; another allows reports to be exported without one. The cross-encoder evaluates the question with each passage and prioritizes the second. The team checks the processed pairs and decides to keep this case distinguishing between operations in the model evaluation.

## How to apply or verify

Record the query-passage pairs sent to the model, its version, and the scores returned. Compare the ranking with human judgments on positive and negative cases. Also check the input limit: a truncated passage may lose exactly the caveat that is needed. Measure the time per [candidate set](/en/resources/glossary/candidate-set) and preserve the results to compare future configuration changes.

## Important distinctions

A cross-encoder is not an answer generator, and its score does not have to be a calibrated probability. The scale depends on the model and the task. Unlike a [bi-encoder](/en/resources/glossary/bi-encoder), joint evaluation does not reuse a single document embedding for all queries. It also cannot find a missing passage on its own: in [reranking](/en/resources/glossary/reranking), it works with the candidates it actually receives.

## 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

- [Bi-encoder](</en/resources/glossary/bi-encoder>)
- [Candidate set](</en/resources/glossary/candidate-set>)
- [Reranking](</en/resources/glossary/reranking>)

## 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>)