AWS Publishes Vector Store Selection Guide

Amazon Web Services has released a comprehensive comparison of vector store backends for Amazon Bedrock Knowledge Bases, targeting developers building retrieval-augmented generation (RAG) applications. The guidance focuses on the customer-managed configuration path, where users select their own vector database rather than using the fully managed option. For additional detail, see AWS Machine Learning Blog.

The analysis examines three supported backends: Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector, and Amazon S3 Vectors. Each option is evaluated across distinct RAG use cases with accompanying benchmarks, though specific numerical results were not disclosed in the announcement. For additional detail, see Amazon Bedrock Knowledge Bases.

Three Vector Store Architectures Compared

Amazon OpenSearch Service provides high-speed results from in-memory data, supporting both managed cluster and serverless deployment models. The service features k-nearest neighbor (k-NN) search and hybrid search that combines lexical and vector approaches. Bedrock Knowledge Bases supports both OpenSearch Managed Clusters and OpenSearch Serverless as backends. For additional detail, see Amazon OpenSearch Service.

Amazon Aurora PostgreSQL with pgvector integrates the relational database capabilities of Aurora with pgvector's similarity search functionality. It supports multiple indexing methods including IVFFlat and HNSW, various distance metrics such as L2, cosine, and inner product, and handles vectors up to 2,000 dimensions in single precision.

Amazon S3 Vectors adds native vector support to the AWS object storage service, designed for cost-effective storage and querying of embeddings at scale. AWS states the service delivers sub-second query performance for similarity searches while reducing vector storage costs by up to 90 percent compared to traditional vector databases.

Ecommerce Product Search Use Case

For ecommerce platforms requiring natural language product search across large catalogs with high concurrent query volumes, AWS recommends Amazon OpenSearch Serverless. The service's hybrid search capabilities combine semantic understanding with traditional keyword matching, which the company says is valuable for product catalog scenarios.

OpenSearch Serverless handles vector search with query latency in the low milliseconds range. Built-in support for complex filtering and aggregations powers faceted navigation such as filtering by price, brand, or color. Multiple distance metrics including cosine similarity and Euclidean distance allow fine-tuning of product similarity calculations.

OpenSearch Performance Optimization Details

The AWS post details several configuration options for OpenSearch Serverless Classic collections that affect vector index performance and cost. Vector embedding size presents a trade-off: larger vectors contain more semantic information but increase resource consumption.

Compression and quantization options are available to balance cost and search quality. The benchmarks referenced in the guidance were conducted on Serverless Classic collections, which require explicit configuration of engine, mode, and HNSW parameters. Managed Clusters offer additional tuning options including auto-optimize, GPU-accelerated indexing, and configurable instance sizing that may produce different results.

NextGen Collections and Availability Notes

Amazon OpenSearch Serverless NextGen collections became generally available in May 2026 but are not yet compatible with the Amazon Bedrock Knowledge Bases Retrieve API. NextGen simplifies index creation by removing engine and mode parameters from index mappings, defaults to 32× compression with GPU-accelerated index builds, and supports scale-to-zero functionality.

For broader guidance across all AWS vector solutions, the company references additional resources including "AWS vector solutions: Build agentic AI where your data lives," "The role of vector datastores in generative AI applications," and "Choosing an AWS vector database for RAG use cases."

RAG Architecture Context

The guidance situates vector stores within the broader RAG architecture, where documents are chunked, embedded, and stored during ingestion. At query time, user queries are converted to vector embeddings, similar chunks are retrieved from the vector database, and passed to a large language model as context for response generation.

The vector database serves as the bridge between raw information and contextual understanding, transforming unstructured data into a searchable, semantically meaningful knowledge space. Vector indexes enable rapid, high-dimensional semantic searches and near-instantaneous retrieval of semantically similar information.

Sources

• AWS Machine Learning BlogSelecting a vector store for Amazon Bedrock Knowledge Bases