Turbopuffer

Turbopuffer is a serverless vector and full-text search database purpose-built for AI workloads at scale. Instead of the traditional in-memory index architecture used by pgvector, Pinecone, or Weaviate, Turbopuffer keeps the primary copy of data on object storage (S3-class) and uses a memory + SSD cache tier to serve queries. That design is what lets it advertise roughly an order-of-magnitude cost reduction versus RAM-resident vector DBs while still delivering ~14ms p50 latency on 1024-dim vectors and 90-100% recall.

The service supports pure vector search (ANN), BM25 full-text search, and hybrid ranking that fuses the two, plus rich metadata filtering. Data is organised into namespaces that scale independently, making it a natural fit for multi-tenant RAG apps that need one logical index per customer, document set, or workspace. A copy-on-write branching feature lets you fork a namespace for experiments, evals, or blue/green re-embedding runs without duplicating storage cost.

It is aimed at engineering teams building retrieval-heavy AI products — RAG assistants, semantic code search, agent memory, recommendation systems, log/observability search — who have outgrown Postgres+pgvector but do not want to run their own Milvus or Vespa cluster. Reference customers include Cursor, Notion, Linear, Superhuman, and Anthropic, and the platform is claimed to handle 4T+ documents, 10M+ writes/s and 25k+ QPS in production. Access is via a REST API and typed client SDKs; there is no self-hosted or open-source distribution — the managed service is the product. Enterprise tiers add BYOC, CMEK, SSO, HIPAA BAA and a 99.95% SLA.