a great way to make search fast is to skip work our basic unfiltered order by on tpuf v3 did a full scan. we cut latency by ~95% with a dynamic filter that skips rows that can't possibly enter the result set v3 is within 70ms of prod, will beat it soon turbopuffer.com/v3
turbopuffer
Technology, Information and Internet
{vector, full-text} search on object storage. fast, cheap, trillion scale. powers Anthropic, Harvey, Notion, and more.
About us
turbopuffer is a fast search engine that combines vector and full-text search using object storage, making all your data easily searchable. Using only object storage for state and NVMe SSD with memory cache for compute, turbopuffer scales horizontally to handle billions of documents. The system caches only actively searched data while keeping the rest in low-cost object storage, offering competitive pricing. Cold queries process 1 million vectors in 500ms (p90), while warm queries take just 20ms (p50). This architecture means it's as fast as in-memory search engines when cached, but far cheaper to run. Storing data in cache and object storage costs less than traditional replicated disk systems, even for frequently accessed data.
- Website
-
https://turbopuffer.com
External link for turbopuffer
- Industry
- Technology, Information and Internet
- Company size
- 11-50 employees
- Type
- Privately Held
- Founded
- 2023
Products
![Click here to view turbopuffer]()
turbopuffer
Enterprise Search Software
turbopuffer is a fast search engine that combines vector and full-text search using object storage, making all your data easily searchable. Using only object storage for state and NVMe SSD with memory cache for compute, turbopuffer scales horizontally to handle billions of documents. The system caches only actively searched data while keeping the rest in low-cost object storage, offering competitive pricing. Cold queries for 1 million vectors take p90=444ms, while warm queries are just p50=8ms. This architecture means it's as fast as in-memory search engines when cached, but far cheaper to run.
Employees at turbopuffer
Updates
-
filters on tpuf v3 are now faster than prod (v2) instead of filtering row-at-a-time, v3 will push compute down to storage to find shortcuts. example: if all rows in a block have the same value for x, then x=8 reduces to one check more in the v3 devlog: [tpuf.link/v3/2026-10-08]
-
september changelog tpuf.link/chlog
-
-
tpuf v3's new storage layout will enable batched execution on all query types — including filters, order by, and aggregations in the latest test run, batched reads made ORDER BY up to 40% faster another small step toward fully vectorized puffin' in prod turbopuffer.com/v3
-
-
we implemented batched reads in the v3 full-text engine, leading to a ~9x performance improvement. postings are stored as columns, and we were reading them a row at a time sometimes optimization is not about being clever, but simply avoiding mistakes tpuf.link/v3/2026-10-02
-
since v1, tpuf has keyed every index on the vector address we've stretched this architecture to its limit. to make search faster in every respect, we must redesign our storage engine so the vector index is no longer primary https://lnkd.in/gyy2wsEB
-
we're changing tpuf's storage architecture so we can keep pushing the search frontier the new engine (v3) now passes CI, but we have to make it fast before we ship it to prod we're grinding query plans in public until v3 wins on everything follow along: turbopuffer.com/v3
-
-
moving embedding into turbopuffer lets us parallelize the embedding call with other work in the query plan whole-system puffin' shaves 100+ ms off your semantic searches https://lnkd.in/gJn4SZwR
-
ask Kepler for a public company's revenue growth, and it traces the number to an exact line item in a verified SEC filing to find the filing, Kepler searches on tpuf → 100M+ financial documents → 85ms p90 latency → 10x lower cost than alternatives https://lnkd.in/gmzfgbSb