I recently created a knowledge base in Amazon Bedrock using the fully managed option. Previously, creating a knowledge base required choosing an underlying vector database such as OpenSearch, but the new setup completed without asking me to configure one. Does anyone know what storage and retrieval services are used behind the scenes, and how this option compares with managing an OpenSearch-backed knowledge base in terms of performance, model selection, flexibility, and cost?
3 Answers
My guess is that AWS is hiding a managed vector-storage layer, possibly based on its newer vector capabilities for DynamoDB or a similar internal service. The pricing appears to be closer to an on-demand managed database model than to operating an OpenSearch cluster, but I’d verify the current pricing before assuming it will be cheaper.
The storage layer wouldn’t be my biggest concern. I’d pay more attention to the managed embedding and reranking models, since those can have a larger effect on search quality. AWS may provide default models while allowing some of them to be changed, so check which settings are exposed and benchmark them against the models you currently use.
AWS now documents the fully managed option separately from the configurations where you choose your own vector store. The exact implementation details and trade-offs are best confirmed in the current Bedrock documentation. If you’re unsure about retrieval quality, the practical answer is to run an evaluation using representative questions and compare it with your existing setup.
That’s the main trade-off I see too: less infrastructure to maintain, but potentially fewer choices over indexing and retrieval behavior. Testing with your own documents and expected answers should reveal whether the defaults are good enough.