turbopuffer

turbopuffer is a serverless vector and full-text search database on object storage, for teams running AI search over large multi-tenant data. Billed by usage.

turbopuffer stores vectors and text for search and retrieval, and answers queries over datasets too large to keep in memory. Its state lives in object storage, while query nodes hold hot data on NVMe SSD and in a memory cache. A single namespace can grow to billions of documents with no cluster sized in advance. Simon Hørup Eskildsen and Justine Li started turbopuffer Inc. in 2023, and the Canadian company builds and operates the service itself. Vector indexes are based on SPFresh, a centroid-based approximate nearest-neighbor method. A namespace can hold several vector columns, each with its own dimensions and its own index. Full-text search ranks with BM25, and multi-queries merge text and vector results through server-side reciprocal rank fusion. Attribute filters run inside vector queries, with recall-aware filtering backed by exact metadata indexes. Native embeddings vectorize text when documents are written and when queries run, and native reranking reorders top results with a model through rerank_by. Sharding spreads a large namespace across nodes, branching takes constant-time snapshots, and copy_from_namespace moves full copies between regions and clouds. The service suits engineering teams that put search or retrieval inside an AI product and expect the corpus to keep growing. SaaS companies with many customers give each tenant a namespace of its own, which keeps every customer's data isolated. Teams that want keyword and semantic relevance in one ranked list get both from a single query. Larger organizations can run it single-tenant or inside their own cloud account, with single sign-on, audit logs and customer-managed keys. Hybrid search behind an application's search box, passage retrieval for an AI assistant and search limited to one customer's documents are the jobs it handles. A retrieval pipeline written in LlamaIndex or LangChain can use turbopuffer as its store and send candidates through a reranker from Cohere, Voyage, MixedBread or ZeroEntropy. Writes become visible to consistent reads once acknowledged, so new documents appear in the next consistent query. For a team whose search index has outgrown memory, it trades the work of sizing and tuning a cluster for a service that scales by namespace and shard.