Amazon S3 Vectors
Native vector storage and similarity search built into S3, operating under a dedicated `s3vectors` AWS service namespace with its own IAM/SCP surface. Organized as Vector Buckets holding Vector Indexes; each index fixes an immutable dimension (1–4096) and distance metric (cosine or euclidean). Supports up to 20 trillion vectors per bucket and 2 billion per index as of late 2025, integrates directly with Bedrock Knowledge Bases and SageMaker Unified Studio, and exports to OpenSearch for hybrid keyword + semantic workloads.
Summary
Native vector storage and similarity search built into S3, operating under a dedicated `s3vectors` AWS service namespace with its own IAM/SCP surface. Organized as Vector Buckets holding Vector Indexes; each index fixes an immutable dimension (1–4096) and distance metric (cosine or euclidean). Supports up to 20 trillion vectors per bucket and 2 billion per index as of late 2025, integrates directly with Bedrock Knowledge Bases and SageMaker Unified Studio, and exports to OpenSearch for hybrid keyword + semantic workloads.
S3 Vectors collapses the "raw data in S3, embeddings in Pinecone" split that most RAG systems ship with. Writes are strongly consistent (a new embedding is immediately searchable), metadata pre-filtering narrows the search space before the distance computation, and queries land around 100ms hot / sub-second cold. The trade: you lose fine-grained ANN knobs — S3 Vectors deliberately hides the underlying HNSW/IVF mechanics and rebuilds indexes autonomously on write.
- Not a full vector database replacement for all use cases. Optimized for scale and cost, not ultra-low-latency real-time search where in-memory indexes win.
dimensionanddistance-metricare set at index creation and cannot change. Plan embedding-model swaps as parallel indexes, not in-place mutations.- The 2-billion-per-index limit means very large deployments still shard across multiple indexes inside a bucket — but bucket-level federation is eliminated.
- "Abstracted ANN" cuts both ways: no HNSW tuning parameters to twist, but also no way to bias a specific workload toward recall or latency beyond the metric choice and filter design.
- Dimension cap is 4,096 — cutting-edge 8K-dim embedding models won't fit. Plan around current generation embedders, not the next one.
- 100 results per query is a hard ceiling — there's no parameter to lift it. Workloads needing top-1000 recall must federate at the application layer.
- No
count(*)aggregation API — counting vectors requires paginating throughListVectors, which is impractical at billion-vector scale. Track counts in your own metadata if you need them. - "$0.05/GB-month" is the storage line; the per-query and per-write API charges are the operational variable. A heavy-read RAG workload at $2.50 per 1M queries can rival the storage line as scale grows.
Pricing posture (April 2026): ~$0.05/GB-month storage + ~$2.50 per 1M queries + ~$0.50 per 1M writes. Worked example: 1M vectors at 1,536 dims ≈ $0.30/month storage. Pinecone Serverless / Weaviate Cloud / Qdrant Cloud start at 1.5–2× the storage price and 3–4× the query price for comparable scale.
The two-tier pattern AWS now recommends: Pair S3 Vectors with Amazon OpenSearch — S3 Vectors as the cost-optimized scalable tier for bulk training data and historical archives, OpenSearch as the real-time tier for production inference and interactive search. Mirrors the file/object split that S3 Files + EFS introduced for filesystems. Bedrock Knowledge Bases, SageMaker Unified Studio, and third-party engines like Spice.ai are the on-ramps that wire the S3-Vectors side natively.
Region availability: 14 regions as of April 2026 (expanded from 5 at preview launch). VPC endpoints strongly recommended for production deployments.
enablesSemantic Search — native vector search capability in S3enablesHybrid S3 + Vector Index — embeddings and raw data co-locatedacceleratesRAG over Structured Data — metadata pre-filtering + strong-consistency writes keep RAG freshness tightsolvesHigh Cloud Inference Cost — removes the standalone vector DB line itemscoped_toVector Indexing on Object Storage — S3-native implementationconstrained_byVendor Lock-In — AWS-specific feature
Definition
Native vector storage and similarity search capability built into Amazon S3, operating within a dedicated `s3vectors` AWS service namespace so vector resources can be governed via distinct IAM policies and SCPs. Embeddings are organized as **Vector Buckets** (parent containers) holding one or more **Vector Indexes** (the spatial query structures). Supports up to 20 trillion vectors per bucket and 2 billion per index as of late 2025, effectively removing the need for query federation across smaller indexes.
Traditional vector search requires a dedicated vector database (Pinecone, Milvus). S3 Vectors eliminates this separate layer by integrating vector storage and search directly into S3, reducing infrastructure complexity.
Serverless semantic search over S3-stored data, RAG without a separate vector DB, vector similarity queries at S3 scale, trillion-scale embedding stores without index federation, Bedrock Knowledge Bases and SageMaker Unified Studio backing stores.
Connections 10
Outbound 8
scoped_to2implements1accelerates1solves1constrained_by1Inbound 2
depends_on1enables1Resources 8
S3 Vectors product page describing the first cloud storage with native vector support at scale.
S3 Vectors launch announcement with architecture details and integration patterns for AI applications.
Official S3 Vectors user guide covering vector bucket creation, ingestion, and similarity search operations.
Index-level reference — immutable dimension (1–4096), distance metric (cosine / euclidean), metadata pre-filter rules, and KMS key scoping.
April 2026 GA expansion writeup — covers the 14-region rollout, the explicit two-tier guidance pairing S3 Vectors with OpenSearch, and the pricing matrix ($0.05/GB-mo + $2.50/1M queries + $0.50/1M writes).
Bedrock Knowledge Bases is the managed RAG on-ramp that wires S3 Vectors into a chunk-embed-store-query pipeline without separate ETL. Reading this clarifies how AWS expects 80% of S3 Vectors usage to look in practice.
GA announcement with scale numbers (20T vectors/bucket, 2B/index), Bedrock / SageMaker integration, and OpenSearch export for hybrid search.
Hybrid keyword + semantic search patterns via S3 Vectors + OpenSearch, including when to export vs keep vectors native in S3.