Configure Automated Embedding for Vector Search

What Automated Embedding Does

Without Automated Embedding, your application must:

1. Generate an embedding for each document that you insert or update.

2. Store that embedding alongside the document in your collection.

3. Generate an embedding for every query, using the same model, at query-time.

4. Submit the query embedding to $vectorSearch.

With Automated Embedding, the mongot process handles steps 1, 2, and 3. Your application:

1. Configures a Vector Search index with the autoEmbed field type for the text field.

2. Inserts and updates documents normally. Then, mongot reads the text field, generates the embedding through the configured embedding model, and stores it.

3. Issues $vectorSearch queries with the query text, not a pre-computed embedding. Then, mongot generates the embedding for the query text.

Supported Embedding Providers

Self-managed mongot Automated Embedding integrates with Voyage AI embedding models.

You incur costs for embedding generation. To learn more, see Manage Billing for Automated Embedding.

Prerequisites

• mongot 1.70.1 or later in your Kubernetes deployment.

• MongoDB 8.2 or later.

• Two Voyage AI API keys.

  Use two separate keys, one for index-time embedding generation and one for query-time. Separate keys isolate query workload from indexing workload and allow independent rate-limit accounting. To learn more, see Rate Limits.

• Outbound network access from mongot to the embedding endpoint.

  The default endpoint is https://ai.mongodb.com/v1/embeddings, which proxies Voyage AI with MongoDB-managed billing, for keys generated using the Atlas UI. You can also use https://api.voyageai.com/v1/embeddings for direct Voyage AI access if you generated the API key directly from Voyage AI.

Generate Voyage AI API Keys

You can generate Voyage AI API keys through either of the following paths:

• (Recommended) Through your Atlas account. Atlas provides API key management with built-in rate-limit configuration. The keys are not tied to an Atlas cluster. To learn more, see Manage API Keys.

• Through Voyage AI directly at voyageai.com.

Generate two keys and name them clearly, for example, mongot-prod-index and mongot-prod-query. Store the keys in secure secret storage.

Configure mongot for Automated Embedding

Create a Kubernetes Secret with both keys:

kubectl create secret generic voyage-api-keys \
  --from-literal=indexing-key=<your-index-key> \
  --from-literal=query-key=<your-query-key> \
  --namespace=<your-namespace>

Reference it from the MongoDBSearch Custom Resource:

apiVersion: mongodb.com/v1
kind: MongoDBSearch
metadata:
  name: mdbs
spec:
  # ...
  autoEmbedding:
    embeddingModelAPIKeySecret:
      name: voyage-api-keys

Optionally, set spec.autoEmbedding.providerEndpoint to override the embedding endpoint that mongot uses. When you omit this field, mongot uses its built-in default endpoint, https://ai.mongodb.com/v1/embeddings. To learn more, see MongoDBSearch Resource Specification.

Create an Automated Embedding Index

After you configure mongot with your Voyage AI API keys, define a Vector Search index that uses the autoEmbed field type. The index definition specifies which text field mongot embeds, which model it uses, the indexing method, quantization, and other parameters. To learn more about the autoEmbed index definition, see MongoDB Vector Search Index Fields.

kubectl exec --context "${K8S_CTX}" -n "${MDB_NS}" mongodb-tools-pod -- \
mongosh --quiet "${MDB_CONNECTION_STRING}" \
   --eval "use sample_mflix" \
   --eval 'db.movies.createSearchIndex("vector_auto_embed_index", "vectorSearch",
      { "fields": [ {
         "type": "autoEmbed",
         "modality": "text",
         "path": "plot",
         "model": "'"${EMBEDDING_MODEL}"'",
         "numDimensions": 1024,
         "similarity": "cosine",
         "quantization": "scalar",
         "indexingMethod": "hnsw",
         "hnswOptions": {
            "maxEdges": 16,
            "numEdgeCandidates": 200
         }
      } ] });'

The following table describes the autoEmbed index definition fields:

You can mix autoEmbed fields with filter fields in the same index. You cannot mix autoEmbed with raw vector fields on the same path.

Run an Automated Embedding Query

Submit query text, not a pre-computed embedding. mongot generates the query embedding for you.

kubectl exec --context "${K8S_CTX}" -n "${MDB_NS}" mongodb-tools-pod -- \
mongosh --quiet "${MDB_CONNECTION_STRING}" \
   --eval "use sample_mflix" \
   --eval 'db.movies.aggregate([ {
      $vectorSearch: {
         index: "vector_auto_embed_index",
         path: "plot",
         query: "spy thriller",
         numCandidates: 150,
         limit: 10
      }
   }, {
      $project: {
         _id: 0,
         plot: 1,
         title: 1,
         score: { $meta: "vectorSearchScore" }
      }
   }]);'

To learn more, see Run Vector Search Queries.

Embedding Storage

mongot persists embeddings in a dedicated internal database on the Kubernetes cluster. mongot does not store embeddings in your source collection. mongot can regenerate embeddings from the source text, and you incur charges for embedding generation.

Changes to the embedding model, output dimensions, or quantization trigger a full re-embedding of the affected index. You incur charges for the regeneration. To learn more, see When mongot Regenerates Embeddings.

To learn more, see Generated Embeddings Collection.

When mongot Regenerates Embeddings

mongot regenerates embeddings for an autoEmbed field when any of the following occur:

• You insert a document.

• You change the embedded field in a document. mongot detects the change through change streams.

• You change the embedding model in the index definition.

• You change the embedding output dimension or data type.

• You change the text field path.

If you change any of the last three items, mongot rebuilds the entire index from scratch. The rebuild can be costly for large collections. Plan model changes deliberately.

Cost and Rate Limiting

MongoDB charges per million tokens at model-specific rates. Indexing cost is proportional to the total text volume of the indexed field across the collection. Query cost is proportional to the volume of query text.

If you use the Voyage AI API key that you created using your Atlas account, you can view API key usage and rate limits in the Atlas UI. To learn more, see Manage Billing for the Embedding and Reranking API. If you use keys created directly from Voyage AI, you must monitor your usage and rate limits in the Voyage AI dashboard.

Voyage AI-side rate-limit errors appear in mongot logs and as failed-document indicators on the affected index. These errors do not disable the index, but they can delay indexing of new documents.

Failure Modes

Limitations

• Preview status. Configuration shape, supported models, and default endpoints might change before Automated Embedding is generally available.

• One embedding model per field. Mixed-model indexes are not supported.

• No embedding model failover. If the Voyage AI endpoint is unreachable, indexing of new documents stalls.