Turbopuffer

Turbopuffer is a managed search engine that runs vector similarity search and BM25 full-text search directly from object storage, an architecture that sets it apart from memory-resident vector databases. Data lives in S3 with a memory and SSD cache layer in front, which keeps costs low for large corpora while still serving warm queries at p50 latencies under 10 milliseconds. Each workload is isolated in its own namespace, and the service scales to billions of vectors with metadata filtering and hybrid queries that combine semantic and keyword relevance. Pricing is usage based rather than instance based, so idle namespaces cost little. It is a closed source commercial service accessed through an HTTP API with official Python and JavaScript clients. Companies including Cursor, Notion, and Linear use it as retrieval infrastructure for AI features, and it suits teams that want production RAG search without operating their own vector database cluster.