Amazon S3 Vectors

AWS lifecycle notice (June 30, 2026) — Amazon Kendra is in maintenance for new customers after July 30, 2026. Evaluate Quick Index or Bedrock Knowledge Bases instead. Full matrix: lifecycle roundup.

Definition

Amazon S3 Vectors is a native vector storage tier on S3 for embeddings and similarity search. Vector buckets store high-dimensional vectors with metadata filters; indexes support cosine, Euclidean, and dot-product distance metrics. S3 Vectors reached GA in 2025 as a Bedrock Knowledge Bases vector store option alongside OpenSearch Serverless, Aurora pgvector, and partner engines — targeting RAG and semantic search where storage cost dominates OpenSearch OCU-hours or dedicated vector DB pods.

As of June 16, 2026, QueryVectors returns up to 10,000 similarity search results per query (100× the prior 100-result limit), with paginated responses via nextToken. Query data-processed charges on indexes with more than 10 million vectors dropped up to 80% automatically. Large result sets may incur data-returned fees beyond the first 512 KB per query — see the S3 pricing page.

The trade-off is latency: expect roughly sub-100ms to low hundreds of ms query times suitable for batch retrieval, wide recall + rerank pipelines, and many chat RAG flows — not sub-10ms agent loops at thousands of QPS.

When to use it

• Bedrock Knowledge Bases RAG with large corpora (10M+ chunks) where OpenSearch baseline OCUs inflate monthly cost — especially after the June 2026 large-index query discount.
• Multi-stage retrieval — wide topK recall, client-side rerank, dedup by document_id — now practical without sharding workarounds for the old 100-result cap.
• Multi-tenant SaaS needing S3-native isolation (prefix or bucket per tenant) with metadata filters at retrieval.
• Archival or long-tail knowledge sets queried occasionally but stored durably for compliance.

When not to use it

• Agentic workflows requiring sub-50ms retrieval inside tight tool-call loops at high QPS — OpenSearch Serverless or in-memory caches win.
• Defaulting to topK=10,000 for simple chat RAG — five chunks to the LLM does not need wide recall; you pay latency and data-returned fees for no gain.
• Hybrid lexical + vector search as a single managed engine — OpenSearch hybrid or Kendra may fit better.
• Graph-heavy relationship traversal — Neptune Analytics combines graph and vector where edges matter.

Tips

• Design metadata fields for mandatory filters (tenant, ACL, doc version) before first ingest — re-indexing billion-vector buckets is painful.
• On wide recall passes, set returnMetadata=True and returnData=False; fetch chunk text only for post-rerank top-N.
• Paginate QueryVectors with nextToken — process the first page while fetching the next; do not buffer thousands of payloads in Lambda memory.
• Upgrade AWS SDKs after June 16, 2026 for pagination support on QueryVectors.
• Run recall@k benchmarks before raising topK; cheapest store is worthless if reranked quality does not improve.

Gotchas

• Serious: Raising topK to thousands with returnData=True without pagination — OOM in Lambda and unexpected data-returned charges past the 512 KB free tier.
• Serious: Using S3 Vectors for real-time agent tool retrieval without load testing — tail latency spikes under concurrent sessions frustrate users.
• Serious: Stale embeddings when source documents change but sync jobs fail silently — pair with document version metadata and health alarms on sync lag.
• Regular: Assuming hybrid keyword search exists natively — you may still need OpenSearch or Athena on structured fields for keyword-heavy queries.
• Regular: Cross-region inference in Bedrock reading vectors in another region adds data transfer — colocate vector buckets with Knowledge Base and model region.

Official references

• Querying vectors — QueryVectors, filters, recall testing.
• Create a vector index — index types and limits.
• Knowledge Bases data source sync — supported ingestion paths.

Related FactualMinds content

• Amazon S3 Vectors: 10,000 Results per Query (June 2026)
• Amazon Bedrock Consulting
• Generative AI on AWS
• Amazon S3
• Generative AI RAG on Bedrock Pattern