AWS Launches Native Vector Search for DynamoDB

AWS has integrated native vector search into Amazon DynamoDB, enabling developers to store embeddings alongside standard application data. The update introduces a new vector index type that supports approximate nearest-neighbor queries directly through a new SearchVectors API. This eliminates the need to synchronize data with external vector databases, simplifying the architecture for applications like retrieval-augmented generation, recommendation engines, and agentic memory.

The serverless feature scales horizontally with no storage limits and supports up to 4096 dimensions. It accommodates Euclidean, Cosine, and Dot product distance functions, alongside inline filtering. Developers can generate embeddings using any model, including Amazon Bedrock Titan Text Embeddings, Cohere Embed, or OpenAI text embedding models. According to AWS executives, the system can scale to trillions of vectors while maintaining single-digit millisecond latency.

Pricing for the new vector index is metered per byte and billed per GB across three dimensions: data written to the index, data processed during searches, and data stored. These fees are added to standard DynamoDB table charges. To manage these costs, AWS architects recommend using lower dimensions, utilizing minimal index projections, excluding embeddings from search results, and employing selective partitioning.

The capability is currently available in all AWS regions supporting DynamoDB for tables using either the Standard or Standard-IA table classes. For local development and self-managed deployments, the team plans to support vector search through ExtendDB, a DynamoDB-compatible adapter.