turbopuffer

turbopuffer is a serverless vector and full-text search database built from first principles on object storage, combining speed, cost-efficiency, and massive scalability. It uses a tiered storage engine with NVMe SSD and memory cache for hot data while keeping the rest in low-cost object storage (S3), delivering sub-10ms p50 warm query latency. Designed for AI applications, semantic search, and recommendation systems, turbopuffer handles billions of documents at a fraction of the cost of traditional vector databases. It is trusted in production at scale, handling 2.5T+ documents, 10M+ writes/s, and 10k+ queries/s.

• Serverless Architecture: Sign up and start indexing immediately — no infrastructure to provision or manage, with automatic horizontal scaling.
• Vector Search: Perform approximate nearest neighbor (ANN) search with 90–100% recall@10, supporting up to 500M documents per namespace.
• Full-Text Search: Built-in full-text search capabilities alongside vector search for hybrid retrieval pipelines.
• Hybrid Search: Combine vector and full-text search in a single query to maximize retrieval quality for AI and LLM applications.
• Metadata Filtering: Filter search results by arbitrary document attributes at query time without sacrificing performance.
• Extreme Cost Efficiency: Usage-based pricing with storage, writes, and queries billed separately — up to 10x cheaper than traditional vector databases.
• Multi-Tenancy & Namespaces: Organize data into isolated namespaces with support for 100M+ namespaces in production.
• Security & Compliance: SOC2 report, GDPR-ready DPA, HIPAA-ready BAA (Scale+), SSO, CMEK, and private networking (Enterprise).
• Simple API: Interact via a straightforward HTTP API with client libraries; get started with the quickstart guide in the docs.
• Benchmarked Performance: Cold queries for 1M vectors at p90=444ms; warm queries at p50=8ms — as fast as in-memory engines when cached.