On August 5, 2026, AWS announced the general availability of vector search within Amazon DynamoDB, enabling vector embeddings to be stored alongside operational data and similarity searches to be performed directly within the table, without copying data to a separate vector database.
The feature delivers single-digit millisecond latency with recall exceeding 99% and is designed to scale to any size, including trillions of vectors. It also relies on DynamoDB’s serverless architecture, so users do not need to provision or patch servers or install and manage software. AWS states that the service requires no version upgrades or maintenance windows, with zero-downtime maintenance.
Reducing the Complexity of AI Application Architectures
Previously, adding vector search to an application using DynamoDB typically required copying data to a dedicated vector store and then creating a pipeline to synchronize the two services. According to AWS, this approach added operational overhead and data-transfer and licensing costs, along with the challenge of maintaining low and predictable latency as the system scaled.
With vector search integrated into DynamoDB, operational data and embeddings can share the same serverless architecture and pay-per-request pricing model. Vector indexes provide unlimited storage capacity, with horizontal scaling as data grows.
The feature targets agent memory applications, retrieval-augmented generation, recommendation engines, personalization, anomaly detection, and other use cases that require semantic retrieval.
How Vector Search Works
The feature creates a new type of index on the attribute that stores vector embeddings. These embeddings can be created using a model selected by the user, such as Amazon Bedrock Titan Text Embeddings, Cohere Embed, or OpenAI text embedding models, and then stored as a list of decimal numbers through the standard PutItem call.
The user then creates a vector index specifying the number of dimensions, the distance function, and any non-vector attributes that can be used to filter results during the search. The SearchVectors API accepts the query vector and the number of requested results, up to a maximum of 100 results, along with optional filtering conditions, and returns the results ranked by similarity.
The feature supports up to 4096 dimensions and the Euclidean, Cosine, and Dot product distance functions, along with inline filtering. AWS recommends matching the distance function with the function used to train the embedding model to achieve the best accuracy.
Example of a Sporting Goods Store
The AWS guide demonstrates adding vector search to an existing product catalog table. In the example, embeddings of product descriptions are stored in a new attribute named descriptionEmbedding, using DynamoDB’s existing List type, without requiring a new data type or a schema change.
After creating an index named ProductDescriptionIndex, marketplace can be used as a partition key to distribute vectors across partitions and restrict the search to a specific marketplace, while category can also be added as a filtering attribute. Filtering conditions support exact matches only and do not support range conditions such as BETWEEN or BEGINS_WITH.
When searching for a natural-language phrase such as “lightweight running shoes for summer,” the user creates a query vector with the same model, then specifies the index, the number of results, the marketplace value, such as US, and the requested category. DynamoDB returns the five most similar products along with their operational attributes, such as name and price, in the same response. With the Cosine and Euclidean functions, lower values indicate greater similarity, while higher values indicate greater similarity when using Dot product.
Availability
Vector search in DynamoDB is now generally available in all AWS commercial Regions, including AWS GovCloud (US) Regions. AWS directs users to the AWS capabilities by Region page for regional availability and the future roadmap, and to the Amazon DynamoDB pricing page for cost details.