Choosing a Vector Store for Amazon Bedrock Knowledge Bases
AWS details how OpenSearch, Aurora PostgreSQL with pgvector, and S3 Vectors differ for RAG workloads on Bedrock, affecting latency and cost.
Original source published: September 17, 2026
When building Retrieval Augmented Generation (RAG) solutions with Amazon Bedrock Knowledge Bases, the vector store selection impacts both performance and expense. Bedrock provides a fully managed option and a customer‑managed path that supports three back‑ends: Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector, and Amazon S3 Vectors.
Amazon OpenSearch (available as managed clusters or serverless) offers low‑millisecond query latency and hybrid search that combines semantic vectors with traditional keyword filtering, which is valuable for e‑commerce product catalog searches that need faceted navigation. Aurora PostgreSQL with pgvector adds relational database features, supports vectors up to 2,000 dimensions, and provides indexing methods such as IVFFlat and HNSW with distance metrics like L2, cosine, and inner product. Amazon S3 Vectors delivers sub‑second similarity queries while reducing vector storage costs by up to 90 % compared with conventional vector databases.
These distinctions help developers and data scientists choose the backend that best matches their RAG latency, cost, and query‑complexity requirements when using Bedrock Knowledge Bases.