$rerank Aggregation Pipeline Stage

Requirements

To use $rerank:

• Ensure that your cluster is running MongoDB 8.3 or later by selecting Latest version with auto-upgrades in the Atlas UI Cluster Builder page.

• Enable Native Reranking through Project Settings. To learn more, see Enable or Disable Native Reranking.

Limitations

You cannot use the $rerank stage:

• For self-managed or Atlas Local deployments.

• For $rankFusion or $scoreFusion input pipelines.

• For queries on a View.

  If you have a MongoDB Search or MongoDB Vector Search index on a View, query the index by running .aggregate() on the source collection instead of the View.

Behavior

$rerank can appear anywhere in the pipeline. However, MongoDB recommends using $rerank after a stage like $search or $vectorSearch that already returns relevant documents in sorted order.

$rerank reranks and returns the first numDocsToRerank documents that are passed to the stage. If $rerank is the first stage, or prior stages do not return deterministically sorted results, then the documents used for $rerank might change between queries.

$rerank returns an error if any of the fields specified in path are not present in one or more input documents. To mitigate this, use the:

• $set stage to set missing fields to an empty string.

• $match stage to filter out documents where fields are missing.

Query

The query.text field determines how the reranker model scores each document. The specified Voyage reranker model computes a relevance score between the query text and the content of each document at the specified path. For most use cases, set $rerank.query.text to the same or similar text as the query from the preceding $search or $vectorSearch stage.

Scoring

Use the score variable with the $meta expression to retrieve the score of each document in the results of the $rerank stage.

{
  "$addFields": {
    "rerankScore": { "$meta": "score" }
  }
}

The $rerank stage replaces the value of $meta: "score" with a new score. To preserve the value from a preceding stage like $rankFusion, you can project the scores into named fields before the $rerank stage.

{
  "$addFields": {
    "originalRankFusionScore": { "$meta": "score" }
  }
},
{
  "$rerank": {
    ...
  }
},
{
  "$addFields": {
    "rerankScore": { "$meta": "score" }
  }
}

Performance and Accuracy

The $rerank stage can be computationally expensive, depending on the size of the input such as the fields to rerank over, the number of documents to rerank, and the query text. $rerank processes queries and documents jointly at query time. Therefore, $rerank might be slower than relevancy search powered by an index, such as the indexes for $search or $vectorSearch. To balance accuracy and performance, configure numDocsToRerank and path to limit the input size. $rerank does not run Voyage AI reranking models on your Atlas cluster resources.

$rerank is best suited for workloads where retrieval quality is prioritized over ultra-low latency, such as RAG and agentic AI applications. For workloads that require ultra-low latency, MongoDB recommends using an index-powered search stage like $search or $vectorSearch.