Troubleshooting

Cannot use the $vectorSearch stage on cluster

To use the $vectorSearch pipeline stage to query your cluster, your cluster must run MongoDB 7.0.2+. If you invoke $vectorSearch on an incompatible version of MongoDB, you might see the following error:

OperationFailure: $vectorSearch is not allowed with the currentconfiguration. You may need to enable the corresponding featureflag.

To check the MongoDB version of your cluster:

If your cluster runs a version of MongoDB earlier than 6.0.11 or 7.0.2, you must upgrade the MongoDB version of the cluster.

Slow queries

For recommendations on improving query performance, see Benchmark for MongoDB Vector Search.

$vectorSearch returns no results

If $vectorSearch queries return no results, perform the following actions:

• Ensure that you're using the same embedding model for both your data and your query. If you embed your query using a different model than you use to embed your data, $vectorSearch can't identify matches.

• Ensure that your MongoDB Vector Search has finished building. When you create or update a MongoDB Vector Search index, the index the index is in an initial sync state. When it finishes building, you can start querying the data in your collection.

Error during document retrieval when using filtering with LangChain

When you use LangChain to perform RAG with pre-filtering, you might encounter the following error:

Error during the document retrieval or generation process:MongoServerError: PlanExecutor error during aggregation :: causedby :: Path 'field' needs to be indexed as token

If an index on field exists, ensure you have created this index as a MongoDB Vector Search index, not a MongoDB Search index. If no index on field exists, create one. To learn more about implementing RAG with MongoDB Vector Search and LangChain, see Answer Questions on Your Data.

Command not found when creating MongoDB Vector Search index

When you attempt to create a MongoDB Vector Search index programmatically, you might encounter a Command not found error. This error occurs for one of two reasons:

• You run the command against a cluster running a version of MongoDB earlier than 6.0.11 or 7.0.2. In this case, you must upgrade the MongoDB version of the cluster to enable MongoDB Vector Search for the cluster.

• You run the command against a Free cluster (formerly known as M0). In this case, if the cluster is running a compatible MongoDB version, you can create a MongoDB Vector Search index with the Atlas UI.

Unable to filter on a given field

MongoDB Vector Search currently supports filtering only on fields with boolean, date, number, objectId, string, and UUID values. To learn more, see About the filter Type.

$rerank request fails with a memory limit error

The $rerank stage buffers the documents to rerank in memory and enforces the following byte limits on that buffer. It rejects, rather than truncates, any request that exceeds a limit:

• A single $rerank request can buffer up to 30 MiB of document data. If the request exceeds this limit, the stage fails with error code 65190 and a message similar to the following:

  $rerank request exceeds the per-request memory limit: 1000documents totalling 40.00 MB against a 30 MB limit. ReducenumDocsToRerank, or add a $project before $rerank to excludefields that are not reranked.

• All concurrent $rerank stages on a node can buffer up to a combined 128 MiB of document data. If a new $rerank request would exceed this limit, the stage fails with error code 65191 and a message similar to the following:

  $rerank cannot be admitted: concurrent $rerank operations onthis node are using 128.00 MB of the 128 MB limit. Retry, orreduce concurrent $rerank load.

To avoid exceeding the per-request limit, reduce numDocsToRerank or use a $project stage before $rerank to exclude fields that you aren't reranking. If you receive the node-wide limit error, retry the query or reduce concurrent $rerank load.