AWS Adds Vector Search to Managed DynamoDB Database Service

Amazon Web Services (AWS) has added vector search capabilities to its managed DynamoDB database service, enabling organizations to store embeddings and run similarity searches without deploying a separate vector database. The new vector search index supports dimensions, distance functions, and metadata filters, and the SearchVectors API can return up to 100 results ranked by similarity. Colin Lazier, vice president of nonrelational databases at AWS, said the serverless architecture is designed to accommodate bursty AI workloads without requiring teams to predict or provision capacity in advance.

TL;DR — Key Takeaways - AWS has added vector search capabilities to DynamoDB, enabling organizations to store embeddings and run similarity searches without deploying a separate vector database. - The new vector index supports dimensions, distance functions and metadata filters, while the SearchVectors API can return up to 100 results ranked by similarity. - DynamoDB’s serverless architecture is designed to accommodate bursty AI workloads without requiring teams to predict or provision database capacity in advance. Amazon Web Services AWS today made available a managed instance of its serverless DynamoDB database that adds support for vector search capabilities that dynamically scale as needed https://aws.amazon.com/blogs/aws/amazon-dynamodb-now-supports-real-time-vector-search-at-any-scale/ . The primary goal is to make it simpler for IT teams to build and deploy artificial intelligence AI applications that take advantage of vector search capabilities that run single-digit millisecond latency without having to deploy a separate vector database, says Colin Lazier, vice president of nonrelational databases at AWS. At the core of that capability is a new vector search index type that has been added to DynamoDB to store vector embeddings within a NoSQL key-value document database that AWS makes available as a managed service. IT teams, as a result, can add a vector index on any attribute that specifies the number of dimensions, the distance function, and any non-vector attributes they want to use as filters to narrow search results at query time. The SearchVectors application programming interface API then accepts a query vector and the number of results to return, up to a limit of 100, and ranks the results by similarity. That capability is especially critical for AI workloads that tend to be a lot more bursty than traditional workloads, which makes it difficult to predict how much IT infrastructure might need to be consumed at any given time, notes Lazier. That approach also provides the added benefit of being able to rely on a serverless computing framework that eliminates the need to assign a team of IT professionals to provision and manage servers to host a separate database, added Lazier. Instead, IT organizations only need to pay for IT infrastructure as it is being consumed, he adds. “You just pay for the storage,” says Lazier. Ideally, more IT teams would spend more time trying to better understand if they have the right tool and platform for the job at hand, says Lazier. There is a tendency to initially deploy a general-purpose relational database to support any application, only to discover that AI applications running at scale have unique requirements, he adds. Of course, as AI continues to evolve, the percentage of AI applications that need vector search capabilities is only going to increase. Rather than spending time managing databases, organizations should be able to devote more resources to building those applications versus having to hire a database administrator DBA to manage a platform deployed by an IT team, notes Lazier. Additionally, IT teams can also use the same managed service to support other classes of workloads that similarly benefit from a serverless NoSQL database, he adds. Most IT teams are already struggling to manage multiple classes of databases, so any opportunity to reduce the number of them required has the potential to reduce the total cost of IT. The challenge and the opportunity now is to determine how best to manage massive amounts of data that are only going to continue to exponentially increase in the age of AI.