What it is
turbopuffer is a hosted search engine for vectors and text that keeps all its data in object storage such as Amazon S3 and caches the active part on NVMe SSD and in memory. That design makes storage cheap and lets one account hold an unlimited number of namespaces (separate indexes, typically one per customer or user).
It offers vector search with filters, BM25 full-text search, hybrid search, regex search and copy-on-write branching of namespaces. Its vector index (ANN v3, updated May 2026) is built on SPFresh, a centroid-based index, with RaBitQ binary quantisation and full-precision rescoring of a small candidate set.
The company was founded in Ottawa in 2023 by former Shopify engineers Simon Eskildsen and Justine Li. Its site names Anthropic, Notion, Atlassian, Linear, Ramp and The New York Times among customers. It is closed source and runs as a multi-tenant service, single-tenant or in your own cloud on the enterprise plan.
Our review
turbopuffer's case is economics at scale. Object storage is the source of truth and only active data is cached on NVMe and in memory, so huge, mostly idle collections cost little: storage is billed at up to $0.33 per GB, and the company's "10x cheaper" claim is its own. The customer list is the strongest evidence it works, with Anthropic, Cognition, Notion, Atlassian, Linear and Harvey on the site. Notion is the telling one; The Information reported in 2025 that losing Notion pushed Pinecone to explore a sale.
The trade-off is spelled out in its own numbers. On 10 million 1,024-dimension documents, warm queries return in 14 ms at p50 and 27 ms at p99, but a cold namespace takes 874 ms at p50 and 1,686 ms at p99. Writes are durable in object storage before they are acknowledged, at a p50 of 165 ms for a 500 kB batch, and each namespace commits at most one write-log entry per second. That suits per-user or per-customer indexes that warm on first use; it does not suit an app where any query on untouched data must be fast. Pinning, capped at 256 namespaces, and pre-flight warming queries are the workarounds.
The limits are generous. Namespaces are unlimited, with more than 250 million seen in production, a single namespace holds up to 128 billion documents, and vector, BM25, hybrid and regex search share one engine with strong consistency by default. The ANN v3 write-up reports p99 of 200 ms at more than 1,000 queries per second across a 100-billion-vector index.
The weaknesses are lock-in and tier gates. There is no open-source edition and no free tier, and the pricing page gates most security features: SSO and a HIPAA BAA start on Scale at $256 a month, while private networking, per-namespace customer-managed keys, BYOC, single-tenancy and the 99.95% uptime SLA are Enterprise-only, from $4,096 a month plus a 35% premium on usage. For one modest index the entry minimum costs more than the usage; the calculator's 1-million-document example comes to about $7 against a $16 floor.
Where it earns its place
A platform team at a B2B SaaS with 50,000 customer workspaces
Give every workspace its own searchable index of documents
One namespace per workspace, with no cap on namespaces and data billed by logical bytes stored. Active tenants stay warm at 14 ms p50; a tenant returning after a quiet week waits close to a second on the first query unless the app sends a warming query at login.
An engineer building code search for an AI coding agent
Index millions of repositories with vectors and BM25 together
Hybrid and regex search run in the same engine, and each namespace accepts 10,000 writes a second. Batch writes, because each one is billed at a 10 kB minimum and larger batches earn up to a 50% discount.
A CTO at a legal-tech company handling client documents
Pass a security review that requires SSO, a BAA and private networking
Scale at a $256 monthly minimum covers SSO, audit logs and a HIPAA BAA, but PrivateLink, per-namespace customer-managed keys and an uptime SLA need Enterprise from $4,096 a month plus a 35% usage premium.
Before you commit
- Setup
- Create an API key and write documents with an official SDK or the HTTP API; namespaces are created on first write, with no index to size.
- Integrations
- Official SDKs for Python, TypeScript, Go, Java, Ruby and C# plus an HTTP API; copy-on-write branching clones a namespace in constant time for testing and migrations.
- Data and privacy
- SOC 2 Type 2 report and a GDPR DPA on every plan; HIPAA BAA from Scale; data stays in the region you choose, encrypted with AES-256, with per-namespace customer-managed keys, private networking, single-tenancy and BYOC on Enterprise.
- Cost at scale
- Storage is billed at up to $0.33 per GB, so 2 TB of vectors and attributes costs up to about $660 a month before writes and queries; queries are billed on the size of the namespace searched, with 80% and 96% marginal discounts above 32 GB and 128 GB.
- Limits
- 10,000 writes a second and 5,000+ queries a second per namespace; 16 concurrent queries on an unpinned namespace; 256 pinned namespaces; eventual-consistency reads can be up to about an hour stale.
What's good
- Object-storage design keeps large, idle data cheap
- Unlimited namespaces suit one index per tenant
- Vector, BM25, hybrid and regex search in one engine
- Strong consistency by default
- SOC 2 Type 2, with a HIPAA BAA from the Scale tier
What's not
- Cold queries take close to a second at p50
- Writes take 165 ms at p50 and commit once a second per namespace
- Closed source; BYOC and single-tenancy only on Enterprise
- No free tier; SSO needs the $256 Scale minimum
- Enterprise adds a 35% premium on usage
Best for
Teams building multi-tenant AI products with many separate indexes or billions of vectors who want no servers to run.
Alternatives
turbopuffer: common questions
- Is turbopuffer worth it?
- turbopuffer is the vector database to beat for products with millions of separate, mostly idle indexes, as long as the first query on a cold one may take a second. We rate it 4.2 out of 5.
- How much does turbopuffer cost?
- Launch $16/month minimum usage; Scale $256/month minimum (SSO, HIPAA BAA); Enterprise from $4,096/month with a 35% usage premium (BYOC, single-tenant, 99.95% SLA). Storage up to $0.33/GB, writes up to $2/GB, queries up to $1/PB queried. No free tier.
- What are the downsides of turbopuffer?
- Cold queries take close to a second at p50; writes take 165 ms at p50 and commit once a second per namespace; closed source; BYOC and single-tenancy only on Enterprise; no free tier; SSO needs the $256 Scale minimum; enterprise adds a 35% premium on usage.
- Who is turbopuffer best for?
- Teams building multi-tenant AI products with many separate indexes or billions of vectors who want no servers to run.
- How much does turbopuffer cost for a team or in production?
- Storage is billed at up to $0.33 per GB, so 2 TB of vectors and attributes costs up to about $660 a month before writes and queries; queries are billed on the size of the namespace searched, with 80% and 96% marginal discounts above 32 GB and 128 GB.
- Is turbopuffer safe for company data?
- SOC 2 Type 2 report and a GDPR DPA on every plan; HIPAA BAA from Scale; data stays in the region you choose, encrypted with AES-256, with per-namespace customer-managed keys, private networking, single-tenancy and BYOC on Enterprise.
- What are the best turbopuffer alternatives?
- The closest alternatives we have reviewed are Pinecone.
Answers checked Oct 22, 2026.
Make turbopuffer? Show this rating on your site
