Vector and full-text search infrastructure
turbopuffer
turbopuffer builds a vector and full-text search database for AI applications. Developers write documents, embeddings, text and metadata through an API, then retrieve information using semantic similarity, keyword matching or hybrid search with filters. Customers include Anthropic, Atlassian, Cursor and Notion. Its architecture separates compute from storage, keeping data in object storage and bringing it into NVMe or memory caches when queried.
Its native embedding service also lets developers send source documents on writes and search text on queries, with turbopuffer generating the corresponding vectors. This supplies an alternative to providing embeddings through a separate application pipeline while retaining access to stored documents and vectors.
Launched in 2023 by former Shopify engineers Simon Hørup Eskildsen and Justine Li, turbopuffer is a Canadian, remote-first company led by Eskildsen as CEO. Its engineering approach includes a clustering-based vector index inspired by SPFresh, which supports incremental updates without periodic rebuilding. Its native filtering connects attribute indexes to those vector clusters, allowing searches to skip clusters without matching documents and evaluate matching candidates within the remaining clusters. This integrates filtering into retrieval instead of applying it only before or after vector search.
In April 2026, the company reported four trillion documents in production and over 25,000 searches per second. Sacra estimated its annualized revenue run rate at $100 million in March 2026. Separately, turbopuffer secured financing of an undisclosed amount from Lachy Groom and Thrive Capital in 2025.
Explore the recordings
This guide maps turbopuffer’s supplied recording archive, which currently contains one joint talk with Legora. Use the paths below to navigate that recording by engineering question. Performance figures, scale estimates and architecture descriptions reflect what the speakers reported in the recording; they are not independently verified statements about today’s products or deployments.
Start with the migration: project boundaries and tail latency
Watch Connect AI to Billions of Legal Documents for Legora’s progression from shared Elasticsearch to regional deployments, Postgres and then turbopuffer. Jacob Lauritzen describes how hashing projects into roughly 4,000 partitions mixed active and inactive data, causing cache thrashing; he reports search and ingestion P99 increasing from about 100 milliseconds to 20 seconds. The useful thread is the relationship between project boundaries, cache behavior and operational choices. The reported move to one turbopuffer namespace per project restored BM25 and improved latency, cost and operational simplicity in the workloads discussed.
Follow the storage architecture and isolation tradeoffs
Return to the same recording for Simon Eskildsen’s explanation of direct object-storage writes, background indexing and reads across memory, SSD and object storage. This path connects idle-project storage economics with namespace-level buckets and encryption keys. The speakers also discuss selected Legora workloads that retained acceptable performance with the SSD cache disabled. Treat that as a workload-specific observation from the talk, rather than a general performance guarantee or confirmation of current isolation features.
Explore legal research, then the index mechanics
Use the talk’s legal-research discussion and concluding index overview to move beyond project-scoped search. The speakers describe fan-out across jurisdictions, authoritative hierarchies, temporal validity and related regulations, with a corpus described in the recording as approaching ten billion vectors. Follow this with the clustered-tree vector-index discussion and the full-text overview of posting lists, BM25, compression and memory bandwidth. This path links retrieval requirements to storage and indexing mechanics; the corpus estimate is a recorded claim, not a current archive-verified count.
2 talks
Newest first
2 speakers at AIE
Affiliations reflect their AIE appearances, not necessarily current employment.
Company sources · checked 2026-09-22
- turbopuffer - fast search engine built on object storage
- turbopuffer the company
- turbopuffer blog
- Native filtering for high-recall vector search
- Building the database for trillion-scale AI search
- Ex-Shopify engineers raise fresh financing to scale Turbopuffer’s AI search
- Turbopuffer revenue, funding & news
- turbopuffer embedding