Amazon S3 Vectors

Definition

Amazon S3 Vectors is a vector storage service that lives natively inside S3 — purpose-built for retrieval-augmented generation (RAG) and similarity search workloads. Announced in 2025, S3 Vectors targets the same workloads as Pinecone, OpenSearch Serverless, and pgvector, but with up to 90% lower storage cost and tighter integration with Amazon Bedrock Knowledge Bases.

What it provides

• Vector buckets — New bucket type for storing embeddings + metadata
• Index types — Cosine, Euclidean, dot-product distance metrics
• Filtered search — Metadata filters at query time (e.g. tenant_id = "acme" AND region = "us-east-1")
• Bedrock Knowledge Bases integration — Listed as a native vector store option alongside OpenSearch Serverless, Aurora, MongoDB, Pinecone
• Hybrid retrieval — Combine with Athena and S3 Tables for metadata-rich retrieval

When to use S3 Vectors

• RAG with > 10M vectors where storage cost dominates the bill
• Multi-tenant SaaS GenAI that needs per-tenant data isolation in S3 prefixes
• Workloads tolerant of ~100ms query latency — S3 Vectors trades query latency for storage cost

When not to use S3 Vectors

• Sub-50ms latency requirements — OpenSearch Serverless is faster on small indexes
• High QPS (> 1,000 queries/sec) — Validate latency at your load before committing
• Need a graph or hybrid lexical-vector store — Neptune Analytics or OpenSearch hybrid are better fits

Cost comparison (illustrative, May 2026)

Always benchmark on your actual vector dimensionality, QPS, and recall requirements before committing.

Common mistakes

Mistake 1: Using S3 Vectors for latency-critical agentic workflows. Use OpenSearch Serverless or in-memory caches for retrieval steps inside an agent loop.

Mistake 2: Skipping the metadata schema design. S3 Vectors lets you filter on metadata at query time — designing the schema for tenant isolation, document version, and freshness up front saves a re-index later.

Mistake 3: Re-creating an entire vector index instead of using the streaming sync from Bedrock Knowledge Bases. Knowledge Bases incremental sync is the supported path.

Related AWS Services

• Amazon Bedrock Knowledge Bases — Native consumer of S3 Vectors
• Amazon OpenSearch Serverless — Lower-latency alternative
• Amazon Aurora pgvector — RDS-based alternative for OLTP-adjacent workloads
• Amazon Neptune Analytics — Graph + vector for relationship-rich retrieval

Related FactualMinds Content

• Amazon Bedrock Consulting
• Generative AI RAG on Bedrock Pattern