Vector Database on Object Storage

Search every
vector.

Vector and full-text search built on S3, GCS, R2, and MinIO — fast, 10× cheaper than traditional vector databases, and infinitely scalable.

p50 latency

<10ms

Cost reduction

10×

Vectors / namespace

100B+

Recall@10

90%+

Your bucket — your data — no egress to us

Works with your existing storage

S3, GCS, R2, MinIO — already yours.

AWS S3
Google Cloud Storage
Cloudflare R2
Azure Blob Storage
MinIO
Tigris
Wasabi
Capabilities

Every search primitive. One API.

01 — Vector Search

Vector Search

Approximate nearest-neighbour over dense embeddings using HNSW. Sub-10ms p50 latency with 90%+ recall — served directly from your object store.

Learn more →
02 — Full-text Search

Full-text Search

BM25 tokenised full-text search across any stored attribute. The same namespace, the same API — no separate search cluster to operate.

Learn more →
03 — Hybrid Search

Hybrid Search

RRF fusion of vector and full-text scores, combined with structured metadata filters. Precision and recall in a single query.

Learn more →

Built for the most demanding search workloads.

Speed

Sub-10ms p50, at any scale

Hot namespaces are served from memory; cold namespaces are fetched from object storage on demand. You pay for queries, not for idle capacity.

Cost

10× cheaper than in-memory DBs

Object storage costs fractions of a cent per GB. No pinned RAM for cold data. No replication at the database layer — your object store handles that.

Scale

Billions of vectors per namespace

No cluster resizing, no shard rebalancing. Storage scales with your object store — which was already built to handle exabytes.

<10ms

p50 latency

10×

cost reduction

100B+

vectors / namespace

90%+

recall@10

Use cases

What teams build with ManyVector.

Dead-simple API

Upsert, query, done.

No schema management. No index tuning. No cluster to provision. Drop in the client, point it at your bucket, and start searching.

<10ms

p50 query latency

10×

cost reduction vs in-memory

100B+

vectors per namespace

90%+

recall@10 default

How it works

Rethinking the vector database from the storage layer up.

Traditional vector databases pin all data in RAM. ManyVector inverts that — object storage is the source of truth, hot data is cached at the query layer.

01 — Architecture

Object store as the source of truth.

ManyVector stores all data in your own S3, GCS, R2, or MinIO bucket. No separate disk tier, no replication overhead — the object store IS the database. You own the bytes.

AWS S3, GCS, R2, MinIO, Azure Blob
No data ever stored on ManyVector disks
$0.023/GB/mo at object-storage prices
Explore storage backends →
02 — Namespace branching

Fork a namespace in milliseconds.

Copy-on-write namespace branching lets you snapshot a collection, test new embedding models, run evals, and roll back — without duplicating a single vector on disk.

Zero bytes copied on fork — metadata only
Instant rollback to any prior snapshot
Eval branches with isolated write paths
Explore namespaces →
03 — Query API

One API. Every search primitive.

Vector, full-text, hybrid, and metadata filter — all in a single consistent query interface. RRF fusion combines vector and BM25 scores. No schema management, no index tuning.

HNSW vector search with 90%+ recall@10
BM25 full-text over stored attributes
RRF fusion with metadata pre-filtering
Explore the API →
What teams say
“We moved from Pinecone to ManyVector and cut our vector storage costs by 10×. The object store architecture means we pay for what we actually query — not for idle RAM.”

Platform Lead — AI-first startup

Built differently
Your bucket — ManyVector never stores your data on our disks
No RAM tax — cold namespaces cost object-storage prices only
Namespace branching — fork, eval, rollback without copying data
Self-hostable — run the entire stack in your own VPC
Compatible with OpenAI, Cohere, and any embedding model

Free tier

100M vectors free

Start building with no credit card. Bring your own S3 bucket and pay only for queries above the free tier.

Sign up free

Pay as you go

Per query, not per GB

ManyVector charges per query — not for storage. Your data lives in your bucket at object-store prices. No idle RAM cost.

See pricing

Self-hosted

Deploy in your VPC

Run ManyVector entirely within your own infrastructure. Air-gapped operation supported. No internet required after setup.

Talk to us

No per-GB markup — you pay your storage provider directly using your own bucket credentials.

Start searching every vector today.