AI product

turbopuffer

turbopuffer is a hosted vector and full-text search database built on object storage, used in document-ingestion systems to persist indexed data with memory or SSD caching. It supports approximate-nearest-neighbor vector search, full-text search, semantic, hybrid, permissioned, and metadata-filtered search through an API, with namespaces for organizing indexed data. Its native filtering searches only the nearest clusters containing potential matches, combining the speed of post-filtering with recall closer to pre-filtering. The service provides automatic scaling, low-latency search, and S3-compatible object storage for semantic search, recommendations, and other AI applications. Kepler uses turbopuffer for hybrid retrieval combining ANN and BM25, followed by reciprocal and cross-encoder re-ranking, and the service has been used for large multi-tenant corpora serving human and agent search.

Visit site Mentioned in 6 videos ↓

What turbopuffer is used for

5 uses taken from transcripts — each links to the moment in the video.

  • Vector storage service used in the ingest-parse-index pipeline to persist data on disk with a memory layer, handling permissioning and metadata filtering at multi-tenant scale.

  • Vector database with 'native filtering' — searches only nearest clusters containing matches, delivering post-filtering speed with near pre-filtering recall, and handling agent-scale query volume (100+ queries/second)

  • The vector database the Vault switched to for search; its 'native filtering' approach scans only the nearest clusters containing matches, delivering post-filtering speed with near pre-filtering recall, and handled the agents' query scale easily.

  • The vector search service Kepler uses for hybrid search, combining ANN and BM25 in a namespace with reciprocal and cross-encoder re-ranking to retrieve and re-rank documents across the citation corpus.

  • A fast and scalable search engine built on object storage that supports keyword, regex, and vector search, designed to solve the challenge of searching large datasets while respecting granular access permissions.

Videos mentioning turbopuffer

6 in the library.