🇬🇧 EN

📊 Reranker

Takes a list of search results and a query, then re-ranks them using a cross-encoder model for improved precision. Use after a RAG Query or Web Search node to surface the most relevant results at the top.

Category: Data Retrieval & Processing · Type identifier: reranker

Overview

Standard vector search and keyword search rank results by similarity scores that are computed independently for each document, without considering the relationship between the query and each result holistically. A cross-encoder reranker solves this by evaluating the query and each candidate result together, producing a more accurate relevance score.

The Reranker node accepts the output of any node that produces an array of results — such as a RAG Query or Web Search node — and returns the same results re-ordered by relevance score. You can optionally limit the output to the top N results, which is useful for controlling how much context you pass to a downstream Agent node.

The recommended pattern is: RAG Query → Reranker → Agent. The RAG Query retrieves a broad set of candidate chunks; the Reranker promotes the most relevant ones; the Agent receives a tighter, higher-quality context window.

Configuration

Field	Status	Description
Results	Required	Reference to a prior step's array of results, e.g. `{{ rag_step.output.results }}`. Each item must have a `content` or `text` field.
Query	Required	The search query used to rank results. Usually the same query passed to the preceding search step.
Provider	Required	The AI provider to use for reranking. Must support reranking models (e.g. Cohere, OpenAI).
Model	Required	The cross-encoder model identifier, e.g. `rerank-english-v3.0`.
Top K	Optional	Return only the top N results after reranking. Default: 5. Leave blank to return all items.

Output Data

The node returns the input results re-ordered by relevance, with a score field added to each item.

Field	Type	Description
`results`	array	Re-ranked array of result objects, each with the original fields plus a `score` property
`results[*].score`	number	Relevance score (0–1, higher is more relevant)
`results[*].content`	string	The result text
`count`	number	Number of results returned

{{ reranker_step.output.results }} {{ reranker_step.output.results[0].score }} {{ reranker_step.output.results[0].content }} {{ reranker_step.output.count }}

Example Usage

RAG → Reranker → Agent, ensuring the agent sees the most relevant chunks first.

Add a RAG Query node. Set Top K to 20 to retrieve a broad candidate set. Use the step key rag_query.
Add a Reranker node. Set Results to {{ rag_query.output.results }}, Query to the same query variable, Provider to Cohere, Model to rerank-english-v3.0, and Top K to 5.
Add an Agent node. Pass {{ reranker_step.output.results }} as the context. The agent now receives only the 5 most relevant chunks.

Tips & Notes

Reranking costs additional API calls. Each result in the input array requires a forward pass through the cross-encoder. Keep the input size reasonable (20–50 results is typical) to control cost and latency.
Input results must have a text field. Each item in the results array must contain either a content or text field. If your upstream step uses a different field name, use a Set Variable node to reshape the array first.
Use Top K to limit agent context. Passing fewer, higher-quality results to an Agent is almost always better than passing many lower-quality ones. A Top K of 3–5 is a good starting point.
Provider must support reranking. Not all AI providers offer reranking models. Check your provider's documentation to confirm support before configuring this node.

Related Nodes

RAG Query — retrieve candidate results from a knowledge base to feed into the Reranker.
Web Search — retrieve live web results to feed into the Reranker.
Agent — consume the re-ranked results as context.