Turbopuffer

Overview

Turbopuffer is a serverless vector and full-text search database built from first principles on object storage, providing fast, cost-effective, and extremely scalable search capabilities.

Performance

• Cold Queries: p90=444ms for 1 million vectors (reads from object storage directly)
• Warm Queries: p50=8ms for 1 million cached vectors
• Sub-10ms p50 latency: When data is cached
• Scale: Hosts 2.5T+ documents, handles 10M+ writes/s, serves 10k+ queries/s

Key Features

• Serverless Architecture: No infrastructure to manage
• Object Storage-First: Built on S3 or GCS for cost efficiency
• Hybrid Search: Vector search and BM25 full-text search
• Automatic Tuning: 90-100% recall without manual optimization
• Clustered Indexes: Uses clustered indexes rather than graph-based approaches (HNSW)
• SSD Caching: Frequently accessed data cached for ultra-low latency

Cost Efficiency

Turbopuffer provides 10x cost savings compared to traditional vector databases:

• Object storage (S3/GCS): ~$0.02 per GB
• SSD caching: ~$0.1 per GB for hot data
• Pay only for what you use, no infrastructure overhead

Architecture

Unlike graph-based approaches that require many round trips, Turbopuffer uses clustered indexes optimized for object storage, reducing latency and costs.

Notable Users

• Notion AI: Powers AI-driven search and recommendations
• Linear: Issue search functionality
• Superhuman: Email search capabilities
• Telus: Enterprise AI copilot
• Cursor: Repository indexing (95% cost reduction)

Use Cases

• Large-scale semantic search
• RAG (Retrieval-Augmented Generation) systems
• Recommendation engines
• Enterprise search applications
• Code repository indexing
• Email and document search
• Customer support knowledge bases

Pricing

Serverless pricing based on:

• Storage usage
• Query volume
• Data transfer

10x more cost-effective than traditional vector database solutions.