What is turbopuffer?
turbopuffer is a managed, serverless search engine that keeps vector and full-text (BM25) indexes on object storage such as S3, caching active data on SSD and in memory so queries stay fast while storage stays cheap.
Checked
Most vector databases keep indexes on servers with local memory and disk. turbopuffer puts the source of truth in object storage and treats SSD and memory as a cache. Storage costs drop towards object-storage prices, and the trade is a slower first query on data that is not yet cached.
That shape suits workloads with a very large total corpus but a small active slice at any moment, such as one index per customer or per code repository. The vendor's homepage names Anthropic, Cognition, Notion, Atlassian, Ramp, Linear, Harvey and The New York Times among its customers.
How it works
Write path
Each write appends a file to the namespace's write-ahead log on object storage and is durable once acknowledged. turbopuffer quotes p50 write latency of 165 ms for 500 kB, and a namespace commits at most one log entry per second, batching concurrent writes.
Indexing
Committed data is indexed asynchronously. Vectors use a centroid-based approximate nearest neighbour index based on SPFresh, chosen to cut round trips to object storage compared with graph indexes such as HNSW or DiskANN. Full-text uses an inverted BM25 index and filters use exact indexes. Unindexed data stays searchable through a slower scan.
Read path
A cold query reads object storage directly; turbopuffer's docs give p50 of 874 ms for 1M documents. Once cached on NVMe SSD, the same query is p50 of 14 ms. A pre-flight query can warm a namespace before latency matters.
Namespaces and consistency
Each namespace has its own prefix and log, and large ones are sharded internally. Queries are strongly consistent by default; eventual consistency lowers warm latency but can be stale by up to about an hour in the worst case.
Pricing and status, checked 1 November 2026
Plans
Launch has a $16 monthly minimum. Scale has a $256 minimum and adds a HIPAA BAA, SSO, audit logs and IP allowlisting. Enterprise starts at $4,096 a month plus a 35% usage premium, and adds single tenancy, BYOC, private networking, customer-managed keys and a 99.95% uptime SLA.
Usage billing
Billed on logical bytes stored, data queried and returned, and writes, with discounts for batched writes. Late-interaction vector queries are billed as two queries.
Recent releases
FTS v2 (4 December 2025), ANN v3 (21 January 2026, which turbopuffer reports at 200 ms p99 over 100B vectors), and native embedding inside turbopuffer (29 September 2026).
Model
Managed only and closed source. There is no self-hosted edition; BYOC is an Enterprise feature.
When to pick it, and when not to
Pick it for
Large or fast-growing corpora where storage cost dominates, many-tenant products that want one namespace per customer, and agent search over code or documents that mixes vectors, BM25 and filters.
Look elsewhere when
Every query must be fast from a cold start, the corpus is small enough to sit in Postgres with pgvector, or policy requires self-hosting without an Enterprise contract.
Compare with S3 Vectors when
The workload is on AWS, queries are infrequent and keyword search is not needed. S3 Vectors has no monthly minimum but no BM25.
Gotchas from production
Cold namespaces
The first query on an uncached namespace takes hundreds of milliseconds. Products with many rarely used tenants should warm a namespace when a user session starts.
Writes batch per second
With one log commit per second per namespace, chatty single-row writes add latency. Batch writes, which also cost less.
Choose consistency per query
Eventual consistency is faster but can return results up to about an hour stale in the worst case, which matters for agents that read back what they just wrote.
Usage bills follow query volume
Charges grow with data queried and returned, so agent loops that search many times per task should be measured before launch, not estimated.
How Motiontic approaches it
Retrieval quality is decided by chunking, hybrid search and reranking more than by the store. Motiontic's engineers measure recall on the client's own documents and questions first, then pick the store whose cost and latency profile fits the measured query pattern.
turbopuffer and its nearest alternatives, checked 1 November 2026
| turbopuffer | Pinecone | Amazon S3 Vectors | |
|---|---|---|---|
| Architecture | Object storage with SSD and memory cache | Managed serverless vector database | Vector buckets and indexes inside S3 |
| Search types | Vector, BM25 full-text, filters, hybrid | Dense, sparse and full-text (BM25 GA 9 Sep 2026) | Vector similarity with metadata filters; no BM25 |
| Hosting | Managed; BYOC on Enterprise | Managed; BYOC available | AWS managed service |
| Entry price | $16 a month minimum | Free Starter; Standard from $50 a month | Pay per use, no minimum |
| Open source | No | No (core engine) | No |
| Multi-tenancy | Namespace per tenant | Namespace per tenant (100 to 1,000,000 per index by plan) | Up to 10,000 indexes per bucket |
Common questions
Is turbopuffer open source?
No. turbopuffer is a managed, closed-source service. Enterprise customers can run it in their own cloud account through BYOC.
How much does turbopuffer cost?
Plans have monthly minimums of $16 (Launch), $256 (Scale) and $4,096 plus a 35% usage premium (Enterprise). Usage is billed on data stored, data queried and returned, and writes.
Does turbopuffer support hybrid search?
Yes. It has BM25 full-text search alongside vector search and metadata filters, and queries can combine them.
Why is the first turbopuffer query slower?
Uncached data is read from object storage. turbopuffer's docs give p50 of 874 ms cold and 14 ms warm for 1M documents, so warming a namespace before a session starts removes most of the gap.
Related
Vector database
Storage and nearest-neighbour search for embeddings, now in many shapes.
Amazon S3 Vectors
Vector storage and similarity search built into S3, priced like storage rather than a database.
Hybrid search
Keyword ranking and vector similarity, fused into one result list.
pgvector vs Pinecone
A Postgres extension compared with a managed serverless vector database.
AI product development
New products and major AI features, built from scratch alongside the client's team.
Sources
- turbopuffer homepage (customers, scale claims)
- turbopuffer pricing (plans and minimums)
- turbopuffer docs, Architecture (write path, cold and warm latency, consistency)
- turbopuffer, ANN v3 (2026-01-21)
- turbopuffer, FTS v2 (2025-12-04)
- turbopuffer, Native embedding (2026-09-29)
- Pinecone pricing
- Pinecone, Full-text search generally available (2026-09-09)
- Amazon S3 pricing, S3 Vectors section
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.