Introduction

                        ╔═ turbopuffer ════════════════════════════╗
╔════════════╗          ║                                          ║░
║            ║░         ║  ┏━━━━━━━━━━━━━━━┓     ┏━━━━━━━━━━━━━━┓  ║░
║   client   ║░───API──▶║  ┃    Memory/    ┃────▶┃    Object    ┃  ║░
║            ║░         ║  ┃   SSD Cache   ┃     ┃ Storage (S3) ┃  ║░
╚════════════╝░         ║  ┗━━━━━━━━━━━━━━━┛     ┗━━━━━━━━━━━━━━┛  ║░
 ░░░░░░░░░░░░░░         ║                                          ║░
                        ╚══════════════════════════════════════════╝░
                         ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

      ╔════════════╗
      ║   client   ║░
      ╚════════════╝░
       ░░░░░║░░░░░░░░
            ▼
╔═ turbopuffer ════════════╗
║  ┏━━━━━━━━━━━━━━━━━━━━┓  ║░
║  ┃    Memory/SSD      ┃  ║░
║  ┃      Cache         ┃  ║░
║  ┗━━━━━━━━┳━━━━━━━━━━━┛  ║░
║           ▼              ║░
║  ┏━━━━━━━━━━━━━━━━━━━━┓  ║░
║  ┃    Object Storage  ┃  ║░
║  ┃      (S3)          ┃  ║░
║  ┗━━━━━━━━━━━━━━━━━━━━┛  ║░
╚══════════════════════════╝░
 ░░░░░░░░░░░░░░░░░░░░░░░░░░░░

turbopuffer is a fast, object-storage native search engine. We service 4T+ documents, 10M+ writes/s, and 25k+ queries/s.

Using only object storage for state and NVMe SSD with memory cache for compute, turbopuffer scales low-latency queries to petabyte scale. This makes turbopuffer as fast as in-memory search engines when cached, but far cheaper to run.

turbopuffer has many features you'd expect from a database optimized for search, including:

• Filters, against an inverted index
• Text search, ranking and boosting
• Vector search, >90% recall, combined with any filters, dense & sparse
• Regex search, with fast trigram regex indexes
• Branching, copy-on-write namespaces
• Encryption, your key or your customer's
• Multi-tenant (default), single-tenant, or BYOC deployments

turbopuffer's tradeoffs for its excellent economics and scalability are higher write latency (p90=248ms for 512KB upserts) from writing directly to object storage, and occasional cold queries for uncached or unpinned data (p90=1214ms on 1M documents). These are excellent tradeoffs for search.

Using object storage as the sole source of truth enables operations like branching — a copy-on-write clone of any namespace, created in constant time regardless of size, with fully independent reads and writes afterward.

turbopuffer is currently focused on first-stage retrieval to efficiently narrow millions of documents (trillions of tokens) down to tens or hundreds documents (thousands of tokens). While it may have fewer features than traditional search engines, this streamlined approach enables higher quality, more maintainable search applications that you can customize in your preferred programming language. See Hybrid Search to get started.

To get started with turbopuffer, see the quickstart guide.