Selecting a vector store for Amazon Bedrock Knowledge Bases

This article compares three vector store backends for Amazon Bedrock Knowledge Bases—Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector, and Amazon S3 Vectors—across distinct RAG use cases with performance benchmarks and selection guidance.

• Amazon OpenSearch Serverless excels at real-time product search with sub-millisecond latency and hybrid search capabilities; benchmarks show 512-dim embeddings match 1024-dim quality at half the size.
• Amazon S3 Vectors provides cost-effective storage for deep research agents, scaling to millions of vectors with sub-linear latency growth and 90% cost reduction versus traditional databases.
• Aurora PostgreSQL with pgvector balances performance and relational data needs for customer-facing chatbots, supporting both HNSW and IVFFlat indexing with sub-50ms query times.
• Selection framework: choose OpenSearch for <50ms latency with hybrid search, Aurora for ~100ms with relational data, S3 Vectors for cost-efficient massive scale.
• Tiered architecture pattern uses S3 Vectors for full corpus storage while promoting high-traffic vectors to OpenSearch or Aurora for performance-critical queries.

Selecting the right vector store requires profiling application latency tolerance, query patterns, and cost sensitivity, then matching to the appropriate backend or combining multiple services in a tiered approach.