MongoDB moves embedding models inside the database: pick one at index time, and vectors follow writes, updates and queries
Synopsis
MongoDB has released Automated Embedding in MongoDB Vector Search as a public preview in Community Edition: a developer chooses an embedding model while creating a vector search index, and text vectors are generated and kept in sync during initial data sync, document inserts and updates, and query processing. Previously developers called external embedding providers from application code before inserting data or running a vector query, and built custom pipelines to keep source data and vectors aligned on every insert, update or delete, plus an extra API call per query. The feature ships Voyage AI's voyage-4-large, voyage-4, voyage-4-lite and voyage-code-3 models and uses asynchronous dynamic batching that adjusts batch size to document shape.
Interpretation
Semantic search in MongoDB no longer requires generating embeddings outside the database; choosing a model when creating a vector index is enough, and vectors are produced automatically on writes and queries. Previously a developer picked one of more than 100 available embedding models, handled external API authentication, input payload and output parameters, and generated embeddings in application code before inserting data or querying vector search. MongoDB says the feature is in public preview in Community Edition, with MongoDB Atlas and Enterprise Edition Vector Search access coming soon, and lists voyage-4-large, voyage-4, voyage-4-lite and voyage-code-3 as the selectable Voyage AI embedding models.
Vectors and source data are maintained by one system, so inserting, updating or deleting a document needs no separate re-embedding pipeline. Every insert, update or delete previously required keeping source data and vector embeddings in sync, which meant custom pipelines, change-stream monitoring and logic to trigger re-embedding when data changed. The system generates vectors automatically during initial data sync when indexes are created, during document insertions and updates, and during query processing, so embeddings stay synchronized as underlying data changes and stale or mismatched embeddings are avoided.
For large ingestions, embedding requests are organized by asynchronous dynamic batching that targets the maximum tokens per request a model allows, adjusting batch size to document shape and size. Embedding millions of documents previously required teams to build MLOps pipelines with queuing systems, specialized workers and retry logic, add monitoring and alerting, and keep slow or failing embedding endpoints from blocking primary database operations. The feature includes built-in automatic retries and error handling; MongoDB says embedding operations do not block primary database operations and that millions of documents can be ingested without building complex MLOps pipelines or monitoring systems.
Perspective
The feature is available now as a public preview in MongoDB Community Edition, with MongoDB Atlas and Enterprise Edition Vector Search access described as coming soon; it works with the Python, JavaScript, Java, C# and GoLang drivers, and with LangChain, LangGraph and the MongoDB MCP server.
MongoDB says the Voyage AI models deliver industry-leading performance on retrieval benchmarks such as RTEB with multilingual support, but gives no figures; the two preview customers describe results qualitatively, one reporting that a 40-million-document corpus turned weeks of engineering into a low-maintenance production-grade workflow, the other reporting immediate performance and simplicity gains in an in-flight production-scale system.