turbopuffer
Object-storage-native vector and full-text search for large retrieval workloads.
About turbopuffer
turbopuffer is an object-storage-native search engine for vector, full-text, hybrid, filtered, and regex retrieval. Durable state stays in object storage, while NVMe SSD and memory act as the compute cache. This design targets lower storage costs at petabyte scale without requiring the entire index to remain in memory. The service reports production workloads spanning more than 4 trillion documents, 10 million writes per second, and 25,000 queries per second. Search options include dense and sparse vectors with filters, text ranking and boosting, trigram-backed regex, encryption, and copy-on-write namespace branching. The tradeoff is uneven latency: direct writes to object storage are slower, and uncached or unpinned data can trigger cold queries. turbopuffer fits first-stage retrieval that reduces a large corpus before application-side reranking or custom logic. Buyers should model the monthly minimums and benchmark cold-query behavior against their own access patterns.
Best for
Best use cases
Choose it when object-storage economics matter more than consistently low write latency.
Choose it when one API should cover vector, full-text, filtered, and regex retrieval.
Test carefully when cold-query latency sits on a user-facing critical path.
Compare before choosing
Lower-cost storage architecture than memory-first search systems, with cache-dependent latency.
Broader retrieval modes than a vector-only database, including text and regex search.
Fewer traditional search-engine features by design, with more application-side customization.
Features
Semantic and hybrid retrieval
Build product or knowledge search that ranks by vector similarity and text relevance, then narrows results with indexed attributes.
First-stage AI retrieval
Narrow millions of records to a smaller result set before reranking or retrieval-augmented generation.
Large multi-tenant search
Operate separate namespaces for customers and branch them with copy-on-write semantics for testing or workflows.
Pros
- Object storage as the source of truth can lower storage cost for very large indexes.
- Vector, full-text, filters, and regex queries reduce the need to operate separate search systems.
- Launch includes all database features at a $16 monthly minimum, making functional evaluation relatively accessible.
Cons
- Direct object-storage writes have higher latency; the documentation reports p90 of 248 ms for 512 KB upserts.
- Uncached or unpinned data can cause cold queries; the documented example reports p90 of 1,214 ms on 1 million documents.
- Scale starts at a $256 monthly minimum, while Enterprise starts at at least $4,096 monthly plus a 35% usage premium.
Pricing Plans
Launch
- All database features
- Multi-tenancy
- SOC 2 report and GDPR-ready DPA
- Community Slack and email support
Scale
- All database features
- Multi-tenancy
- HIPAA-ready BAA
- Single Sign-On
- Audit logs
- IP allowlisting
- Private Slack channel
- 8–5 support hours
Enterprise
- All database features
- Multi-tenancy, single-tenancy, or BYOC
- Private networking
- Customer-managed encryption per namespace
- 24/7 support
- Support SLA
- 99.95% uptime SLA
* Pricing information is for reference only. Please refer to the latest announcement on turbopuffer's official website.