I created a knowledge base in Amazon Bedrock and noticed that it can now be created without separately selecting a vector store such as OpenSearch. Previously, configuring an underlying vector database appeared to be required. Does anyone know what storage and retrieval infrastructure the fully managed option uses, how it compares with an OpenSearch-backed knowledge base, and whether there are differences in cost, embedding models, ranking models, or retrieval quality?
3 Answers
The database layer would not be my main concern. The bigger question is which embedding and reranking models the managed configuration uses. It may default to Amazon-provided models, but Bedrock appears to allow choosing different models in some configurations. Check the model settings and run retrieval evaluations before deciding whether the defaults are good enough for your data.
There is a comparison in the Bedrock documentation that explains the available managed and custom storage options. If you are unsure about quality or cost, benchmark the managed knowledge base against your existing OpenSearch setup using representative queries, measuring answer accuracy, retrieval relevance, latency, and total cost. There is no reliable substitute for evaluating it with your own content.
The fully managed option likely hides the underlying storage and retrieval components rather than eliminating them. Amazon may be using a managed service built around its own vector-search infrastructure, potentially including DynamoDB-backed vector support. The exact implementation is not completely exposed, so the best way to compare it with OpenSearch is to test both with the same documents and questions.
That seems plausible. Vector support was added to DynamoDB around the same time, and the pricing appears similar to DynamoDB on-demand pricing, although the managed knowledge base may include additional service charges.