turbopuffer reimagines vector database architecture by building directly on top of object storage rather than using traditional database storage engines. This fundamental design choice eliminates the provisioned compute and storage costs that make conventional vector databases expensive at scale — customers pay only for the storage their data consumes and the compute their queries use, with automatic scaling that handles traffic spikes without manual capacity planning. The result is vector search that costs roughly one-tenth of equivalent deployments on Pinecone, Weaviate, or Qdrant, making it economically viable to index and search billions of embeddings.
The platform combines vector similarity search with full-text BM25 search in a single query interface, enabling hybrid retrieval strategies that use both semantic and keyword matching. This eliminates the common pattern of running separate vector and text search systems and merging results at the application layer. Queries support metadata filtering with arbitrary predicates, allowing precise retrieval like finding semantically similar documents that also match specific categories, date ranges, or user permissions. The serverless architecture means indices are always available without cold starts, and write throughput scales automatically as data volumes grow.
turbopuffer's customer roster includes some of the most demanding AI workloads in production: Anthropic uses it for internal retrieval systems, Cursor relies on it for codebase search across millions of repositories, and Notion integrates it for AI-powered document search. The official site now reports 4T+ documents, 10M+ writes/s, and 25k+ queries/s in production systems, a vendor-published scale signal that should be attributed rather than treated as an independent benchmark. Funded by Thrive Capital and Lachy Groom with reported revenue growth of 10x in 2025, turbopuffer represents the serverless, cost-optimized future of vector search infrastructure.
