ES|QL RERANK command

Serverless Preview Stack Preview 9.2.0

The RERANK command uses an inference model to compute a new relevance score for an initial set of documents, directly within your ES|QL queries.

PUT _cluster/settings
{
  "persistent": {
    "esql.command.rerank.limit": 5000
  }
}
		

PUT _cluster/settings
{
  "persistent": {
    "esql.command.rerank.enabled": false
  }
}
		

FROM books
| WHERE title:"search query"
| SORT _score DESC
| LIMIT 100
| RERANK "search query" ON title WITH { "inference_id" : "my_rerank_endpoint" }
		

Syntax

RERANK [column =] query ON field [, field, ...] [WITH { "inference_id" : "my_inference_endpoint" }]
		

Parameters

column

(Optional) The name of the output column containing the reranked scores. If not specified, the results will be stored in a column named _score. If the specified column already exists, it will be overwritten with the new results.

query

The query text used to rerank the documents. This is typically the same query used in the initial search.

field

One or more fields to use for reranking. These fields should contain the text that the reranking model will evaluate.

my_inference_endpoint

The ID of the inference endpoint to use for the task. The inference endpoint must be configured with the rerank task type.

Description

The RERANK command uses an inference model to compute a new relevance score for an initial set of documents, directly within your ES|QL queries.

Typically, you first use a WHERE clause with a function like MATCH to retrieve an initial set of documents. This set is often sorted by _score and reduced to the top results (for example, 100) using LIMIT. The RERANK command then processes this smaller, refined subset, which is a good balance between performance and accuracy.

Requirements

To use this command, you must deploy your reranking model in Elasticsearch as an inference endpoint with the task type rerank.

RERANK commands may time out when processing large datasets or complex queries. The default timeout is 10 minutes, but you can increase this limit if necessary.

How you increase the timeout depends on your deployment type:

If you don't want to increase the timeout limit, try the following:

• Reduce data volume with LIMIT or more selective filters before the RERANK command
• Split complex operations into multiple simpler queries
• Configure your HTTP client's response timeout (Refer to HTTP client configuration)

Examples

Rerank search results using a simple query and a single field:

FROM books METADATA _score
| WHERE MATCH(description, "hobbit")
| SORT _score DESC
| LIMIT 100
| RERANK "hobbit" ON description WITH { "inference_id" : "test_reranker" }
| LIMIT 3
| KEEP title, _score
		

Rerank search results using a query and multiple fields, and store the new score in a column named rerank_score:

FROM books METADATA _score
| WHERE MATCH(description, "hobbit") OR MATCH(author, "Tolkien")
| SORT _score DESC
| LIMIT 100
| RERANK rerank_score = "hobbit" ON description, author WITH { "inference_id" : "test_reranker" }
| SORT rerank_score
| LIMIT 3
| KEEP title, _score, rerank_score
		

Combine the original score with the reranked score:

FROM books METADATA _score
| WHERE MATCH(description, "hobbit") OR MATCH(author, "Tolkien")
| SORT _score DESC
| LIMIT 100
| RERANK rerank_score = "hobbit" ON description, author WITH { "inference_id" : "test_reranker" }
| EVAL original_score = _score, _score = rerank_score + original_score
| SORT _score
| LIMIT 3
| KEEP title, original_score, rerank_score, _score