turbopuffer
Free tierFast vector and full-text search engine built on object storage — 10x cheaper and infinitely scalable
Free tier available·Technical·API available
Key strengths
10x cheaper than traditional vector databases due to object storage architectureSub-10ms p50 latency with Memory/SSD caching layerMassive scale: 4T+ documents, 10M+ writes/s, 25k+ queries/s in productionHybrid search combining vector (ANN) and full-text (BM25) in one systemUnlimited namespaces with instant copy-on-write branching
Free tier + paid plans
San Francisco, USA
Founded 2023
No ratings yet
Technical Documentation
turbopuffer exposes a straightforward HTTP API with the following core primitives:
- Write — Upsert or delete documents (vectors + metadata attributes) into a namespace. Supports batch writes up to 10k writes/s per namespace (32 MB/s).
- Query — Execute ANN vector search, BM25 full-text search, or hybrid search with metadata pre/post-filters and custom ranking. Returns top-k results with optional recall tuning.
- Namespace metadata — Retrieve stats and configuration for a given namespace.
- Authentication — All requests require a bearer token passed via standard HTTP
Authorizationheader.
Architecture: Writes are durably persisted to object storage (S3). A Memory/SSD cache layer is maintained per namespace for warm queries (p50 ~14ms on 10M × 1024-dim vectors). Cold namespaces hydrate on first access. Namespace branching provides instant copy-on-write snapshots for isolation, testing, or parallel workloads.
Limits (current production):
- Max documents per namespace: 500M @ 2TB
- Max namespaces: Unlimited
- Max global write throughput: Unlimited
- Vector search recall@10: 90–100%
SDKs and quickstart guides are available in the docs at /docs.