turbopuffer

turbopuffer is a search database designed for vector and full-text search, built entirely on object storage such as Amazon S3. It operates as a fast, scalable, and cost-effective alternative to traditional search engines or specialized vector databases. The core architecture separates compute from storage: a client layer handles query requests, which are processed through a memory and SSD cache before reaching the underlying object storage. This design allows turbopuffer to achieve low latency while keeping storage costs roughly 10 times lower than conventional solutions, as object storage is inherently cheaper than managing dedicated SSD or RAM clusters. The service supports two primary search modalities: vector search, which is essential for semantic similarity matching in AI applications like recommendation systems, retrieval-augmented generation (RAG), or embedding-based clustering; and full-text search, which provides exact keyword matching, BM25 scoring, and boolean queries for traditional search use cases. Both modes can be combined in hybrid queries to balance relevance and recall. Data is indexed and stored in object storage, with caching layers (memory and SSD) accelerating frequently accessed queries. The system automatically scales from zero to huge data volumes without manual provisioning, because object storage capacity is virtually unlimited and cache layers can be added or removed dynamically. A notable feature is branching, which enables instant, copy-on-write namespaces. This lets users create lightweight snapshots or experimental branches of their search index for testing, A/B comparisons, or versioned deployments without duplicating underlying data. Branches share the same storage until modified, making them fast and storage-efficient. This is partic