85 agentic tools audited for security incidents · Talk to the lab →
MotionticBook a call
Tool explainer

What is Amazon S3 Vectors?

Amazon S3 Vectors is an AWS storage service that adds vector buckets and vector indexes to S3, so embeddings can be stored and queried by similarity, with metadata filters, without running a vector database.

Checked

S3 Vectors treats embeddings as another kind of S3 data. A team creates a vector bucket, adds indexes, writes vectors with metadata, and queries for the nearest matches. There are no clusters to size and no servers to patch, and the price is set by storage and use rather than provisioned capacity.

It entered preview in July 2025 and became generally available on 2 December 2025. AWS positions it as the cheap, durable tier for large or rarely queried corpora, with Amazon OpenSearch Service as the tier to move vectors into when they need high-throughput, low-latency search.

How it works

  • Buckets, indexes and vectors

    A vector bucket holds up to 10,000 indexes, and each index holds up to 2 billion vectors of 1 to 4,096 dimensions, each with a key and metadata.

  • Queries

    QueryVectors returns the nearest neighbours to a query vector, filtered on metadata. AWS gives under one second for infrequent queries and about 100 ms for frequent ones.

  • Index modes

    Since 30 September 2026, ENHANCED indexes resolve the metadata filter first and then search only the matching vectors. CLASSIC indexes filter during the search. Buckets created from that date default to ENHANCED.

  • Integrations

    A vector store for Amazon Bedrock Knowledge Bases, a storage layer for Amazon OpenSearch Service, plus CloudFormation, PrivateLink and resource tags.

Limits and pricing, checked 1 November 2026

  • Scale

    Up to 2 billion vectors per index, 10,000 indexes per bucket and 10,000 vector buckets per Region per account. Available in 31 AWS Regions since 31 March 2026.

  • Queries

    Up to 10,000 results per query since 16 June 2026 (previously 100), returned in pages of up to 100.

  • Writes

    Up to 1,000 combined PutVectors and DeleteVectors requests a second per index, 2,500 vectors inserted or deleted a second, and 500 vectors per call.

  • Metadata

    Up to 40 KB per vector in total, of which 2 KB can be filterable, and up to 50 metadata keys.

  • Price in the N. Virginia Region

    $0.06 per GB-month of storage, $0.20 per GB of vector data uploaded, $2.50 per million query calls, plus a per-TB charge on data processed that falls as the index grows. There is no monthly minimum.

When to pick it, and when not to

  • Pick it for

    Large, cold or per-tenant corpora on AWS, archives of embeddings kept for later reuse, and Bedrock Knowledge Bases where a team wants the lowest-maintenance store.

  • Look elsewhere when

    Search needs BM25 or hybrid ranking, answers must come back in tens of milliseconds every time, or the index takes a constant stream of updates.

  • Use it as a tier when

    A hot subset belongs in OpenSearch or another engine and the long tail can sit in S3 Vectors, which AWS supports directly.

Gotchas from production

  • Filtered queries on CLASSIC indexes

    Filtering during the search can return fewer results than asked for; AWS's own example asked for 10 tenant-scoped matches and got 2. Switching an index to ENHANCED applies in place, with no re-ingestion and no extra charge.

  • Old buckets keep the old mode

    Indexes in buckets created before 30 September 2026 stay on CLASSIC until updated, so an audit of index modes belongs on the upgrade list.

  • No keyword search

    Exact terms such as part numbers, names and error codes are where vector-only retrieval misses. Pair S3 Vectors with a BM25 engine when those queries matter.

  • Write throughput is per index

    Streaming single-vector updates are capped per index, so bulk loads should batch up to 500 vectors a call and spread across indexes where the design allows.

  • Large top-k costs more

    Raising top-k towards 10,000 returns more data, which adds latency and cost. Most agent retrieval needs far fewer results, followed by a reranker.

How Motiontic approaches it

Retrieval quality is decided by chunking, hybrid search and reranking more than by the store. Motiontic's engineers measure recall on the client's own documents and questions first, then pick the store whose cost and latency profile fits the measured query pattern.

Amazon S3 Vectors and its nearest alternatives, checked 1 November 2026

Amazon S3 Vectors and its nearest alternatives, checked 1 November 2026
Amazon S3 VectorsturbopufferPinecone
What it isVector buckets and indexes inside S3Search engine on object storage with SSD cacheManaged serverless vector database
Search typesVector similarity with metadata filtersVector, BM25 full-text, filters, hybridDense, sparse and full-text
Pricing modelStorage, uploads and queries; no minimumUsage with $16 to $4,096 monthly minimumsFree Starter; Standard from $50 a month
Storage price$0.06 per GB-month (N. Virginia)Billed on logical bytes stored$0.33 per GB-month on Standard
Multi-tenancyIndex per tenant, up to 10,000 per bucketNamespace per tenantNamespace per tenant (100 to 1,000,000 per index by plan)
Where it runsAWS only, 31 Regionsturbopuffer cloud; BYOC on EnterprisePinecone cloud; BYOC available

Common questions

Is Amazon S3 Vectors a vector database?

AWS describes it as object storage with native support to store and query vectors, not a database server. It covers similarity search with metadata filters, without keyword search or provisioned capacity.

How much does S3 Vectors cost?

In the N. Virginia Region: $0.06 per GB-month of storage, $0.20 per GB uploaded and $2.50 per million queries, plus a per-TB charge on data processed by each query. There is no monthly minimum.

How many results can an S3 Vectors query return?

Up to 10,000 since 16 June 2026, delivered in pages of up to 100. Before that the limit was 100 per query.

Can S3 Vectors be used with Bedrock Knowledge Bases?

Yes. S3 Vectors is a supported vector store for Amazon Bedrock Knowledge Bases, and it can also serve as the storage layer for Amazon OpenSearch Service.

Running agents that have to hold up?

Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.