turbopuffer
@turbopuffer.bsky.social
{vector, full-text} search engine built on object storage. fast, cheap, trillion scale. powers Anthropic, Harvey, Notion, Cognition, and more
Mem0 migrated 400M+ agent memories from pgvector to turbopuffer, solving a semi-selective filtering problem that spiked tail latencies in Postgres → 150k+ isolated search indexes → 70ms p90 hybrid retrieval latency → 97% vector recall@10 tpuf.link/mem0
Mem0 migrates 400M+ agent memories from pgvector to turbopuffer
Postgres didn't scale for Mem0's agent memory platform. They migrated hundreds of millions of memories to turbopuffer, reducing end-to-end latency by 70x.
tpuf.link
tpuf now supports late interaction [beta] use models like ColBERT to represent text as a set of vectors (1 per token) tpuf uses a single-vector ANN index for a fast first pass, then reranks hits using exact late interaction scoring to boost recall docs: turbopuffer.com/docs/query#...
autoscaling is deceptively hard our indexer fleet scales nodes on job queue time. if a queue suddenly went quiet, we'd scale down too hard & new jobs could queue up waiting for nodes to claim them we tweaked the HPA signal to prevent underprovisioning → ~2x shorter queue time
new: computed attributes compute additional values on the results of the rank_by and pass them as extra signals to a reranker → BM25 score for a vector query → vector distance for a BM25 query → individual scores for composite rank_by clauses docs: turbopuffer.com/docs/query#...
new: highlighting extract the text fragments most relevant to a query → highlight matches in your search results UI → minimize context passed to your LLM docs (and playground): turbopuffer.com/docs/fts#hi...
now in beta: native embeddings in tpuf embedding is the most painful part of puffing. we want to make it easy you can now convert chunks to vectors as you read and write to turbopuffer, without extra calls to an embedding model provider API docs: turbopuffer.com/docs/embedding
puffin' in paris for RAISE summit we're bringing together good friends for a night of champagne, caviar, and chicken nuggets july 9. rsvp here: luma.com/raisesummit...
Legora searches 2B+ legal documents on turbopuffer → strict per-matter data isolation with CMEK → 10x lower tail latency than Postgres → 98% avg recall@10 tpuf.link/legora
Legora searches 2B+ legal docs on turbopuffer
Legora replaced Elasticsearch and pgvector + DiskANN with turbopuffer to build an efficient retrieval layer that minimizes input tokens without sacrificing performance, all with legal-grade data isolation and security.
tpuf.link
new: i8 vectors f32: 4 bytes/dim i8: 1 byte/dim 4x fewer bytes → 75% lower storage and query costs + faster queries when embedded with a quantization-aware model (e.g. voyage-4-large) trained on i8 vectors, recall loss can be ~0! docs: turbopuffer.com/docs/perfor...
we cut our base price from $64 → $16/mo start puffing for 4x less
we open sourced alyze, the Rust crate behind tpuf's default full-text search tokenizer (word_v4) our first tokenizers (up to word_v3) were built on Tantivy's analyzer, and we owe them many thanks alyze does a bit less, but does it up to ~4x faster github.com/turbopuffer...
Atlassian's cross-product AI platform, Rovo, searches 5B+ documents on turbopuffer BYOC → 19% increase in search quality → 60ms p90 latency → 96% average recall@10 tpuf.link/atlassian
Atlassian finds its multi-cloud BYOC search engine in turbopuffer
Atlassian chose turbopuffer for its operational and architectural simplicity, cloud-agnostic BYOC deployment model, low cost, and virtually unlimited scalability. It now underpins search for millions of Atlassian customers on Jira, Confl...
tpuf.link
a year ago, ~98% of tpuf queries were vector ANN last 30d: 64% vector ANN 19% full-text BM25 13% filter-only 3% aggregate 1% other (sparse vector, exact kNN, ...)
new: rerank_by before, you'd implement rank fusion client-side. now, a little QoL upgrade, especially nice for large result sets docs: turbopuffer.com/docs/query#...
new: branching create an instant, copy-on-write clone of a tpuf namespace → constant-time (440ms p50, ~1s p99) → fully independent → unlimited branches, unlimited branch depth docs: turbopuffer.com/docs/branching
tpuf quantizes vectors to improve perf (RaBitQ) the algo randomly rotates vectors, and we were using matmul at O(d²) space & time, brutal at high dims. 10k = 400MB in RAM! we rebuilt the rotation using FWHT at O(d) space & O(d log d) time. ~no recall loss, 10k = only 5kB in RAM
new in turbopufer: the Fuzzy filter typo-tolerant substring matching with a configurable edit distance, so you can puff (or puf) even when you spell it wrong docs: tpuf.link/fuzzy
SID-1 is an agentic search model → 1.9x recall over RAG + rerank → 24x faster, 99% cheaper than GPT-5.1 trained using large-scale RL on turbopuffer at 1k+ QPS bursts over 10M+ document corpora across thousands of steps tpuf.link/sid-1
Training SID-1 to beat GPT-5 at search with 1k+ QPS RL
SID-1 is an agentic search model that is 24x faster than GPT-5.1-high, 374x cheaper than Sonnet 4.5, and achieves 1.9x higher recall than traditional RAG pipelines. Here's how we trained it using large-scale RL on turbopuffer.
tpuf.link
filtered counts on tpuf just got much faster our ANN index may replicate docs across clusters for better recall, so we had to dedupe matching IDs (slow) now, we store a bitmap of replica positions so the query plan is pure bitmap ops: (filter_bitmap - replica_bitmap).popcnt()
we're bringing more of the database to the tpuf dashboard step one: namespace metadata more soon
new: sparse vectors a first-class retrieval primitive that composes with BM25 + attribute ranking in the same query plan, no client-side fusion needed for SPLADE / learned-sparse retrievers (or roll your own weights for custom feature scoring) docs: turbopuffer.com/docs/query#...
new: namespace pinning pin namespaces to reserved compute for faster, cheaper, more predictable p99 on high QPS workloads ~50x cheaper than shared compute for a 128GB namespace at 500 QPS docs: turbopuffer.com/docs/pinning
new tpuf regions, so you can puff a little closer to home 🇧🇷 aws sa-east-1 (São Paulo) 🇺🇸 gcp us-east1 (South Carolina) 🇧🇪 gcp europe-west1 (Belgium)
✨ docs search for search docs ✨ now live on turbopuffer.com/docs
BM25 efficiently scores text, but relevance often depends on more than text (recency, popularity, PageRank) we score non-text attributes as clauses in the same MAXSCORE plan that evaluates BM25 → better first-stage relevance, still scales to 100M+ tpuf.link/rank-by-attr
Mixing non-text attributes into text search for better first-stage relevance
turbopuffer now allows you to combine attribute values into the scoring function of text queries. Ranking by attribute helps achieve better relevance in the first-stage with the same scalability characteristics as BM25.
tpuf.link
stemming is what makes a text search for "run" match documents containing "running" or "runs" we just shipped a small stem cache so repeated terms skip the stemmer → ~2x tokenization throughput when stemming is enabled
