Amazon Web Services has made vector search generally available in DynamoDB across all commercial AWS regions, including AWS GovCloud.
The launch adds native vector search to the managed NoSQL database, letting customers store vector embeddings in the same tables as operational data and run similarity searches without moving data to a separate vector database.
The change targets a common setup in generative AI and recommendation workloads, where companies often use one database for live application data and another specialised system for semantic search. The new approach removes the need to replicate data and maintain a synchronisation pipeline between services, AWS said.
Vector search has become a standard building block for applications that use large language models and other AI systems because it helps software retrieve semantically similar records instead of relying on exact keyword matches. In practice, that supports uses including retrieval-augmented generation, agent memory, recommendations, personalised search and anomaly detection.
DynamoDB users can add a vector index to an attribute that stores embeddings and query that index through the SearchVectors application programming interface, according to AWS. The service returns up to 100 results ranked by similarity, with optional filters on non-vector attributes.
DynamoDB supports embeddings of up to 4,096 dimensions and three distance functions: Euclidean, Cosine and Dot product. It also allows inline filtering on selected attributes, although filter conditions are limited to exact matches rather than range-based queries.
How it works
Under the new model, a developer generates embeddings using a model of their choice and stores them in a DynamoDB table as a list of floating-point numbers. They then create a vector index on that attribute and define parameters including dimension count, distance function and any filterable attributes.
AWS illustrated the feature with an online sporting goods store. In that example, product descriptions are converted into embeddings and added to an existing product catalogue table, allowing a shopper to search with a phrase such as lightweight running shoes for summer instead of exact keywords.
The index can also be scoped with a partition key, allowing searches to run within a defined segment of the data, such as a single national marketplace, rather than across a broader index.
Search results are returned with the operational attributes already stored in the table, such as name and price. The meaning of the similarity score depends on the distance function used: lower values represent closer matches for Cosine and Euclidean measures, while higher values represent closer matches for Dot product.
Market context
The move reflects a broader push by large cloud providers to fold vector search into existing data platforms rather than leave it solely to specialist vector database vendors. For customers already standardised on DynamoDB, AWS is positioning the feature as a way to keep AI-oriented retrieval within an existing operational data layer.
That may appeal to engineering teams looking to cut data movement and reduce the complexity of running separate stores for transactional and semantic workloads. It also strengthens DynamoDB's position as AWS ties its database portfolio more closely to AI application development.
The service remains serverless and follows the same pay-per-request pricing model as the wider DynamoDB platform, AWS said. Vector indexes also scale horizontally as data volumes grow, including for very large collections of vectors.
The addition is notable because many organisations have treated vector search as a separate architectural layer, often adding operational overhead and software costs. By placing the feature directly in DynamoDB, AWS is betting that much of the demand will come from customers that prefer to extend an established database rather than adopt another specialised system.
DynamoDB has long been used for web-scale application back ends that require low-latency reads and writes. The vector search release expands that role into AI retrieval tasks, particularly where semantic search needs to sit close to user, product or event data already stored in the service.
AWS said the feature is designed for deployments ranging up to trillions of vectors, with support for single-digit millisecond latency at more than 99% recall.