# turbopuffer turbopuffer is an object-storage-native search engine for vector, full-text, hybrid, and regex search. It helps AI, search, and data teams retrieve content from large document collections with filtering, ranking, and scalable namespace-based storage. - Findings24: - Site: - Preços: Freemium - Categorias: Motor de Busca IA, API de pesquisa com IA, Banco de dados vetorial ## O que é turbopuffer? turbopuffer is a hosted, object-storage-native search engine for vector, full-text, hybrid, and regex search. It uses object storage as the durable source of state, with memory and NVMe SSD caches for query execution, so applications can search large collections without keeping the entire index in memory. The service supports metadata filtering, ranking, boosting, document lookup, and namespace-based organization. It is intended for search across text, images, videos, web pages, agent memories, and other indexed data. The documented API includes operations to create, update, and delete documents, plus queries with filters and ranking. Its architecture is designed for large-scale workloads: the site reports production systems handling more than 1 trillion documents, 10 million writes per second, and 25,000 queries per second. The documented tradeoffs include higher write latency from writing directly to object storage and slower occasional queries for uncached or unpinned data. ## O que turbopuffer pode fazer? ### Vector and hybrid search Run approximate nearest-neighbor vector searches with documented recall above 90%, and combine dense or sparse vector search with full-text strategies. ### Full-text and regex search Search text with an inverted index, BM25 ranking, and boosting. Fast trigram regex indexes support pattern-based retrieval in addition to keyword and semantic search. ### Filtering and ranked queries Apply metadata filters while querying documents, then use ranking and boosting to control how matching results are ordered. ### Object-storage-native scaling Use object storage for durable state and memory or NVMe SSD caches for compute. The architecture is intended to scale across very large document collections and high-throughput workloads. ### Namespaces and branching Organize data in namespaces for multi-tenant or application-specific isolation. Copy-on-write branching creates a namespace clone in constant time while keeping later reads and writes independent. ### Flexible deployment and security options The pricing tiers list multi-tenant deployment across plans, with single-tenant and BYOC options at Enterprise. Available higher-tier controls include SSO, audit logs, private networking, and per-namespace CMEK. ## Casos de uso ### Grounding AI agents Give agents semantic and full-text tools for retrieving relevant documents, web content, memories, reasoning transcripts, tool calls, or evaluation data during a task. ### Search across text and media Index and filter large collections of documents and multimedia files so applications can locate the small set of relevant records, files, or media assets for a user. ### Web and research search Build search tools backed by an index of news, papers, or broader web content, allowing agents and applications to retrieve from a maintained corpus. ### Semantic recommendations Match articles, products, or other content to items a user has liked or engaged with using semantic similarity search. ### Large-scale model data workflows Curate substantial datasets for pre-training or post-training models to search through reinforcement-learning workflows. ## Perguntas frequentes ### What search methods does turbopuffer support? The documented product supports vector search, BM25 full-text search, hybrid search combining vector and full-text strategies, metadata filtering, and regex search using trigram indexes. ### How does turbopuffer store and query data? Object storage is the durable source of state, while memory and NVMe SSD caches provide compute for queries. Cached data can deliver low latency; uncached or unpinned namespaces can produce slower cold queries. ### What are the main architectural tradeoffs? Writing directly to object storage increases write latency; the documentation gives a p90 of 248 ms for 512 KB upserts. Uncached or unpinned data can also have occasional cold-query latency, with a documented p90 of 1,214 ms for a 1-million-document example. ### How is data organized for multi-tenant applications? Applications can separate data into namespaces. The product supports multi-tenant deployment by default on listed plans, while Enterprise also offers single-tenant and BYOC deployment options. Copy-on-write branching can create independent namespace copies. ### Which plans and deployment options are available? The pricing page lists Launch, Scale, and Enterprise plans. Launch has a $16-per-month minimum, Scale has a $256-per-month minimum, and Enterprise starts at $4,096 per month. Enterprise adds deployment and support options such as single-tenant, BYOC, private networking, and service-level commitments. ## Informações rápidas - **Category**: Vector and full-text search database - **Storage model**: Object storage with memory and NVMe SSD caching - **Search modes**: Vector, full-text, hybrid, regex, and filtered queries - **Document organization**: Namespaces with copy-on-write branching - **Documented production scale**: 1T+ documents, 10M+ writes/s, and 25k+ queries/s - **Plans**: Launch, Scale, and Enterprise