AWS has added native vector search to Amazon DynamoDB. Applications can now store vectors alongside their regular data and find similar information without maintaining a separate vector database.
What is vector search?
A vector is a numerical representation of the meaning of text, an image, or other content. This representation allows an application to find related information even when it does not use exactly the same words.
Developers can create a vector index in DynamoDB and use it to retrieve the items most similar to a query. Results are ordered according to their similarity.
AI agent memory and RAG systems
AWS lists several uses for the new feature, including semantic search, RAG systems, personalized recommendations, and memory for artificial intelligence agents.
For example, an agent can store representations of previous conversations, tasks, or results. When it receives a new request, it can retrieve the most relevant information and use it as context.
Data and vectors in one platform
Before this update, an application using DynamoDB might have needed another service to store and search its vectors. This required copying data and keeping both systems synchronized.
With native vector search, regular data and vector representations can remain on the same platform. This reduces the need for a separate vector database when operational data already lives in DynamoDB.
Main limits
The feature supports vectors with up to 4,096 dimensions and can return up to 100 results per search.
It also supports three ways to calculate similarity: cosine, Euclidean distance, and dot product. The appropriate configuration depends on the model used to create the vectors and the needs of each application.
Availability and costs
Vector search is generally available in commercial AWS Regions. Availability of the models used to create vectors may vary by Region.
DynamoDB remains a serverless service, so applications do not need to manage machines. However, vector indexes use additional space and generate storage and usage charges. Creating vectors with an external model may also have a separate cost.
