Turbopuffer

// rankings · cost per standard workload

// overview

turbopuffer is a vector and full-text search database built from first principles on object storage (S3) with NVMe SSD/memory cache. It handles 4T+ documents, 10M+ writes/s and 25k+ queries/s in production, offering sub-10ms p50 latency when cached, >90% recall, automatic scaling and 10x cost savings versus alternatives. Key capabilities include vector search, full-text (BM25), hybrid search, metadata filtering, regex, branching (copy-on-write namespaces) and encryption.

Best for: AI applications, semantic search and recommendation systems needing petabyte-scale similarity search at low cost

// pricing

// security & compliance

// features

• vector search (>90% recall, dense/sparse, any filters)
• full-text search (BM25 ranking/boosting)
• hybrid search
• metadata filtering via inverted index
• branching (instant copy-on-write namespaces)
• regex search with trigram indexes