turbopuffer
turbopuffer is a vector and full-text search database delivered as a managed service. It is uniquely architected on object storage (e.g., Amazon S3) with an in-memory/SSD caching layer, allowing it to serve queries with low latency while leveraging the cost-effectiveness and scalability of cloud object storage. This design enables turbopuffer to be an order of magnitude cheaper than traditional vector databases, while supporting petabyte-scale datasets, up to 4 trillion+ documents, 10 million+ writes per second, and 25,000+ queries per second in production. Key features include sub-10ms p50 query latencies for warm data, automatic scaling, and support for hybrid vector and full-text search over a single dataset. It provides metadata filtering to narrow search results and recently introduced branching, which enables instant copy-on-write namespaces for experimentation and versioning. The service is used by leading technology companies like Cursor, Anthropic, Notion, Atlassian, and Linear for AI-powered search, recommendation systems, and knowledge retrieval. turbopuffer’s cost model is consumption-based, and its performance is optimized for large-scale, high-throughput workloads. The system decouples storage from compute, allowing independent scaling. Cold data may experience slightly higher latency (sub-second), but the caching layer minimizes this for frequently accessed data. The database is accessed via a REST API, and official client libraries are available for easy integration. It is designed for developers building modern AI applications that require fast, affordable, and scalable search over embeddings and text.
- Category
- Vector DB
- Pricing
- Paid
Problem solved
Traditional vector databases often require expensive dedicated infrastructure and struggle to scale cost-effectively to billions of vectors. turbopuffer addresses this by leveraging cheap, durable object storage as the primary storage layer, reducing costs by ~10x compared to alternatives. It provides fast, low-latency search over massive datasets, enabling teams to build AI-powered features without worrying about infrastructure management or runaway expenses.
Use cases
- Semantic search over large document collections (e.g., Notion's AI search)
- AI-powered code search and retrieval in developer tools (e.g., Cursor)
- Hybrid vector + full-text search for issue trackers (e.g., Linear's issue search)
- Recommendation engines requiring low-latency similarity search at scale
- Building retrieval-augmented generation (RAG) pipelines for large language models
- Scaling vector search for enterprise knowledge bases with metadata filtering
Audience
- Developers
- ML Engineers
- Data Scientists
Pros
- 10x cheaper than traditional vector databases
- Sub-10ms p50 latency on warm data
- Scales to trillions of documents and petabyte-size datasets
- Automatic scaling without manual sharding or reindexing
- Hybrid vector and full-text search with metadata filtering
Cons
- Cold data may exhibit higher latency (hundreds of milliseconds)
- Fully managed service, potentially vendor lock-in
- Not open source; cannot be self-hosted
Official website: https://turbopuffer.com/
Canonical LambdaBase page: https://www.lambdabase.com/tools/turbopuffer