turbopuffer
@turbopuffer.bsky.social
search every byte 🔍 {vector, full-text} search engine built from first principles on object storage. 10x cheaper, scales to 100B. powers Notion, Cursor, Linear
February 4, 2026 at 5:00 PM
new: ContainsAnyToken filter
return documents that match any token in the query string
faster than BM25 when you just need a binary match (e.g. "showing 36 of 131,072" in your search UI) as it skips score computation entirely
docs: tpuf.link/containsany
return documents that match any token in the query string
faster than BM25 when you just need a binary match (e.g. "showing 36 of 131,072" in your search UI) as it skips score computation entirely
docs: tpuf.link/containsany
January 28, 2026 at 6:02 PM
new: ContainsAnyToken filter
return documents that match any token in the query string
faster than BM25 when you just need a binary match (e.g. "showing 36 of 131,072" in your search UI) as it skips score computation entirely
docs: tpuf.link/containsany
return documents that match any token in the query string
faster than BM25 when you just need a binary match (e.g. "showing 36 of 131,072" in your search UI) as it skips score computation entirely
docs: tpuf.link/containsany
8.3% of query CPU in our largest (~4 TiB) tpuf namespaces was being spent on hash-map lookups to figure out which files might contain a key range
fix: precompute it in a flat array
trade memory for cache-friendly access → 8% faster puffin'
fix: precompute it in a flat array
trade memory for cache-friendly access → 8% faster puffin'
![Rust code snippet showing before/after. HashMap lookup replaced by precomputed nested array access for faster range retrieval.
before: HashMap lookup at each step in binary search
self.sstables.get(table_id).unwrap().range
after: precomputed in parallel vec
self.level_ranges[i][j][k]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:622wudn2gyeskbjhj75uxglg/bafkreihcyymbswry2ctcoxb4nicx3p5n7jfolstq4dtgwxchskqkalc4di@jpeg)
January 26, 2026 at 3:00 PM
8.3% of query CPU in our largest (~4 TiB) tpuf namespaces was being spent on hash-map lookups to figure out which files might contain a key range
fix: precompute it in a flat array
trade memory for cache-friendly access → 8% faster puffin'
fix: precompute it in a flat array
trade memory for cache-friendly access → 8% faster puffin'
tpuf ANN v3, for when you need to index the entire web
100B+ vectors @ 50ms p50 / 200ms p99 latency
blog: tpuf.link/ann-v3
100B+ vectors @ 50ms p50 / 200ms p99 latency
blog: tpuf.link/ann-v3
ANN v3: 200ms p99 query latency over 100 billion vectors
Our latest ANN release supports scales of 100+ billion vectors in a single search index, with 200ms p99 query latency at 1k QPS and 92% recall.
tpuf.link
January 21, 2026 at 7:52 PM
tpuf ANN v3, for when you need to index the entire web
100B+ vectors @ 50ms p50 / 200ms p99 latency
blog: tpuf.link/ann-v3
100B+ vectors @ 50ms p50 / 200ms p99 latency
blog: tpuf.link/ann-v3
BM25 queries use an inverted index: each search term maps to a posting list of matching document ids
tpuf now dynamically adjusts the encoding of postings lists based on density, like roaring bitmaps
result → up to 26% faster search for queries whose terms match many documents
tpuf now dynamically adjusts the encoding of postings lists based on density, like roaring bitmaps
result → up to 26% faster search for queries whose terms match many documents
January 16, 2026 at 4:00 PM
BM25 queries use an inverted index: each search term maps to a posting list of matching document ids
tpuf now dynamically adjusts the encoding of postings lists based on density, like roaring bitmaps
result → up to 26% faster search for queries whose terms match many documents
tpuf now dynamically adjusts the encoding of postings lists based on density, like roaring bitmaps
result → up to 26% faster search for queries whose terms match many documents
![turbopuffer December 2025 changelog
🧱 Redesigned inverted index structure for faster FTS queries
📤 New object storage-native indexing queue for up to 10x faster queue time
🔦 kNN exact search for 100% recall on filtered vector search queries
🪣 Return a max number of search results per attribute value using limit.per
🇨🇦 AWS ca-central-1 (Montreal) region
🌏 Cross-region backups guide
🤝 Link multiple orgs to a single account for unified billing, SSO, and roles [opt-in, beta]](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:622wudn2gyeskbjhj75uxglg/bafkreigjhxta2yqn4fk35cl4uc6r36blefvyhxca43v77tanvvbzxc32mi@jpeg)
January 15, 2026 at 7:42 PM
for FTS v2, we redesigned our inverted index structure
• tighter compression
• less KV overhead
• better MAXSCORE interaction
up to 10x smaller indexes → up to 20x faster text search!
tpuf.link/fts-index
• tighter compression
• less KV overhead
• better MAXSCORE interaction
up to 10x smaller indexes → up to 20x faster text search!
tpuf.link/fts-index
Designing inverted indexes in a KV-store on object storage
How we redesigned our inverted index structure using fixed-sized posting blocks to achieve 10x smaller indexes and dramatically better throughput.
tpuf.link
January 14, 2026 at 7:45 PM
for FTS v2, we redesigned our inverted index structure
• tighter compression
• less KV overhead
• better MAXSCORE interaction
up to 10x smaller indexes → up to 20x faster text search!
tpuf.link/fts-index
• tighter compression
• less KV overhead
• better MAXSCORE interaction
up to 10x smaller indexes → up to 20x faster text search!
tpuf.link/fts-index
we modeled BM25 query latency across varying term counts, document counts, and top_k values.
takeaways:
- sometimes longer queries are *faster*
- the longer the query, the less efficiently it scales
- fitted curves on raw data are immensely gratifying
tpuf.link/bm25-latency
takeaways:
- sometimes longer queries are *faster*
- the longer the query, the less efficiently it scales
- fitted curves on raw data are immensely gratifying
tpuf.link/bm25-latency
Why BM25 queries with more terms can be faster (and other scaling surprises)
I analyzed how BM25 query latencies scale with document count and top_k. Longer queries scale less efficiently, and essential terms impact performance in some surprising ways.
tpuf.link
January 8, 2026 at 2:54 PM
we modeled BM25 query latency across varying term counts, document counts, and top_k values.
takeaways:
- sometimes longer queries are *faster*
- the longer the query, the less efficiently it scales
- fitted curves on raw data are immensely gratifying
tpuf.link/bm25-latency
takeaways:
- sometimes longer queries are *faster*
- the longer the query, the less efficiently it scales
- fitted curves on raw data are immensely gratifying
tpuf.link/bm25-latency
we rolled out a new indexing queue on all tpuf shared regions
~10x lower index queue time → new documents get indexed sooner → faster queries on new data with less WAL scanning
built entirely on object storage, no kafka
(chart: max create_index time in queue, gcp us-east4)
~10x lower index queue time → new documents get indexed sooner → faster queries on new data with less WAL scanning
built entirely on object storage, no kafka
(chart: max create_index time in queue, gcp us-east4)
December 23, 2025 at 4:35 PM
we rolled out a new indexing queue on all tpuf shared regions
~10x lower index queue time → new documents get indexed sooner → faster queries on new data with less WAL scanning
built entirely on object storage, no kafka
(chart: max create_index time in queue, gcp us-east4)
~10x lower index queue time → new documents get indexed sooner → faster queries on new data with less WAL scanning
built entirely on object storage, no kafka
(chart: max create_index time in queue, gcp us-east4)
TELUS indexes 25,000+ AI copilots on turbopuffer
"This doesn't exist without turbopuffer. If we used a traditional search provider, our bill would be in the millions each year." - Justin Watts, Distinguished Engineer
tpuf.link/telus
"This doesn't exist without turbopuffer. If we used a traditional search provider, our bill would be in the millions each year." - Justin Watts, Distinguished Engineer
tpuf.link/telus
TELUS indexes 25,000+ AI copilots on turbopuffer
serverless vector and full-text search built from first principles on object storage: fast, 10x cheaper, and extremely scalable
tpuf.link
December 22, 2025 at 6:58 PM
TELUS indexes 25,000+ AI copilots on turbopuffer
"This doesn't exist without turbopuffer. If we used a traditional search provider, our bill would be in the millions each year." - Justin Watts, Distinguished Engineer
tpuf.link/telus
"This doesn't exist without turbopuffer. If we used a traditional search provider, our bill would be in the millions each year." - Justin Watts, Distinguished Engineer
tpuf.link/telus
new: phrase matching
use the ContainsTokenSequence in your tpuf FTS queries to match documents on an exact phrase
docs: tpuf.link/phrase-match
use the ContainsTokenSequence in your tpuf FTS queries to match documents on an exact phrase
docs: tpuf.link/phrase-match
December 16, 2025 at 4:00 PM
new: phrase matching
use the ContainsTokenSequence in your tpuf FTS queries to match documents on an exact phrase
docs: tpuf.link/phrase-match
use the ContainsTokenSequence in your tpuf FTS queries to match documents on an exact phrase
docs: tpuf.link/phrase-match
new: Sum aggregate function
combine it with group_by and filters for olap-style puffin'
combine it with group_by and filters for olap-style puffin'
December 15, 2025 at 4:00 PM
new: Sum aggregate function
combine it with group_by and filters for olap-style puffin'
combine it with group_by and filters for olap-style puffin'
yesterday: short human queries, scalar CPUs
today: long LLM queries, wide SIMD lanes
for FTS v2, we use a vectorized MAXSCORE algorithm instead of WAND, because dumb & serial beat smart & random algorithms on modern CPUs
tpuf.link/maxscore
today: long LLM queries, wide SIMD lanes
for FTS v2, we use a vectorized MAXSCORE algorithm instead of WAND, because dumb & serial beat smart & random algorithms on modern CPUs
tpuf.link/maxscore
Vectorized MAXSCORE over WAND: building faster text search for humans and agents
turbopuffer has improved text search performance up to 20x by adopting the same text search algorithm as Apache Lucene, a vectorized variant of block-max MAXSCORE
tpuf.link
December 12, 2025 at 2:00 PM
yesterday: short human queries, scalar CPUs
today: long LLM queries, wide SIMD lanes
for FTS v2, we use a vectorized MAXSCORE algorithm instead of WAND, because dumb & serial beat smart & random algorithms on modern CPUs
tpuf.link/maxscore
today: long LLM queries, wide SIMD lanes
for FTS v2, we use a vectorized MAXSCORE algorithm instead of WAND, because dumb & serial beat smart & random algorithms on modern CPUs
tpuf.link/maxscore
PlayerZero's AI retrieves and reasons over code and logs to debug large codebases.
"We went from 3 minutes to <10 seconds, with much better recall, which means fewer agent loops." - Maria Vinokurskaya, Founding Engineer
tpuf.link/playerzero
"We went from 3 minutes to <10 seconds, with much better recall, which means fewer agent loops." - Maria Vinokurskaya, Founding Engineer
tpuf.link/playerzero
December 10, 2025 at 2:59 PM
PlayerZero's AI retrieves and reasons over code and logs to debug large codebases.
"We went from 3 minutes to <10 seconds, with much better recall, which means fewer agent loops." - Maria Vinokurskaya, Founding Engineer
tpuf.link/playerzero
"We went from 3 minutes to <10 seconds, with much better recall, which means fewer agent loops." - Maria Vinokurskaya, Founding Engineer
tpuf.link/playerzero
turbopuffer queries are strongly consistent, which requires scanning new writes in the WAL while indexing happens async
tpuf's WAL scan is now up to 2x faster
tpuf's WAL scan is now up to 2x faster
December 8, 2025 at 3:55 PM
turbopuffer queries are strongly consistent, which requires scanning new writes in the WAL while indexing happens async
tpuf's WAL scan is now up to 2x faster
tpuf's WAL scan is now up to 2x faster
December 5, 2025 at 4:00 PM
FTS v2: up to 20x faster full-text search
turbopuffer is now on par with Tantivy and Lucene for many queries, more to come
v2 now in beta. 2 PRs away from all query plans being implemented. will be enabled in prod for all, shortly.
turbopuffer is now on par with Tantivy and Lucene for many queries, more to come
v2 now in beta. 2 PRs away from all query plans being implemented. will be enabled in prod for all, shortly.
December 4, 2025 at 6:20 PM
FTS v2: up to 20x faster full-text search
turbopuffer is now on par with Tantivy and Lucene for many queries, more to come
v2 now in beta. 2 PRs away from all query plans being implemented. will be enabled in prod for all, shortly.
turbopuffer is now on par with Tantivy and Lucene for many queries, more to come
v2 now in beta. 2 PRs away from all query plans being implemented. will be enabled in prod for all, shortly.
new: cross-region and cross-org copy_from_namespace
copy your documents wherever you need to puff with a 50% discount - no more restrictions
good for manual backups, testing, and branching
copy your documents wherever you need to puff with a 50% discount - no more restrictions
good for manual backups, testing, and branching
December 3, 2025 at 4:00 PM
new: cross-region and cross-org copy_from_namespace
copy your documents wherever you need to puff with a 50% discount - no more restrictions
good for manual backups, testing, and branching
copy your documents wherever you need to puff with a 50% discount - no more restrictions
good for manual backups, testing, and branching
November 14, 2025 at 1:07 PM
tuning a query planner is an evergreen task. this month we improved p99.99 latency by >90% for this order-by query workload.
November 14, 2025 at 12:15 PM
tuning a query planner is an evergreen task. this month we improved p99.99 latency by >90% for this order-by query workload.
a database is nothing without its guarantees
as part of our recent patch_by_filter launch, we clarified the ACID properties users can expect from turbopuffer
read more here: turbopuffer.com/docs/guaran...
as part of our recent patch_by_filter launch, we clarified the ACID properties users can expect from turbopuffer
read more here: turbopuffer.com/docs/guaran...
Guarantees
serverless vector and full-text search built from first principles on object storage: fast, 10x cheaper, and extremely scalable
turbopuffer.com
October 28, 2025 at 3:00 PM
a database is nothing without its guarantees
as part of our recent patch_by_filter launch, we clarified the ACID properties users can expect from turbopuffer
read more here: turbopuffer.com/docs/guaran...
as part of our recent patch_by_filter launch, we clarified the ACID properties users can expect from turbopuffer
read more here: turbopuffer.com/docs/guaran...
new: `patch_by_filter` updates all documents that match the provided filter, making bulk updates a breeze!
October 27, 2025 at 3:00 PM
new: `patch_by_filter` updates all documents that match the provided filter, making bulk updates a breeze!
new: rank by filter! boost scores when docs match a condition (e.g. scale > 1pib). plugs straight into rank_by, and works alongside full-text search
October 24, 2025 at 3:00 PM
new: rank by filter! boost scores when docs match a condition (e.g. scale > 1pib). plugs straight into rank_by, and works alongside full-text search
October 15, 2025 at 4:28 PM
tpuf ANN v3 can search 100 billion vectors with a p99 of 200ms
simplicity scales
[v3 in beta, unfiltered search, 1024D, k=10, 92% recall]
simplicity scales
[v3 in beta, unfiltered search, 1024D, k=10, 92% recall]
October 14, 2025 at 6:59 PM
tpuf ANN v3 can search 100 billion vectors with a p99 of 200ms
simplicity scales
[v3 in beta, unfiltered search, 1024D, k=10, 92% recall]
simplicity scales
[v3 in beta, unfiltered search, 1024D, k=10, 92% recall]