turbopuffer

What it is for

turbopuffer is an object-storage search engine: you upsert vectors and text into a namespace, then query ANN, BM25, or both, while the source of truth stays on S3 (or equivalent) and a memory/SSD cache keeps hot data fast. Use it when the job is cheap, huge search you meter in storage / write / query bytes, not when you need a $20 chat seat.

What makes it unique: the cover is the official turbopuffer/turbopuffer-python GitHub file list — the public Python client, not a pricing card.

Key features

• Vector ANN — dense (and sparse) nearest-neighbor search with published recall@10 in the 90–100% band (docs; homepage limits). One similarity query. Dimension count and f16 vs f32 are SKU of this vector.
• Full-text search — lexical ranking and boosting on indexed attributes, not an embedding (docs “Text search”). A keyword query. Regex / trigram is a mode of this inverted index, not a fifth product.
• Namespaces — the isolation unit (filters, tenants, per-namespace throughput). Multi-tenant is the default; single-tenant and BYOC are deploy SKUs of this.
• Copy-on-write branching — clone a namespace in constant time; the branch then has independent reads and writes (docs “Branching”). A snapshot, not a query.
• Object-storage source of truth — writes go to object storage; memory/SSD is a cache (docs architecture). The cost model, not an API. CMEK and Private Networking are Enterprise add-ons.

Hybrid “vector + FTS in one request” is the same two indexes composed, not a third engine. Attribute filters attach to this query. The Python / JS clients are install. Storage, writes, and queries are the meter.

Who it's for

Best fit: search / RAG teams who will pay the Launch minimum of $16/month and grow on logical GB, not a per-seat vector UI.

Pricing

Verified 2026-10-22 (Asia/Shanghai) from turbopuffer.com/pricing. Commercial-only. Dollars are usage + a plan minimum. Unused commitment does not roll over.

• Launch — minimum $16/month. All database features. Multi-tenancy. SOC 2 report + GDPR-ready DPA. Community Slack & email.
• Scale — usage plus the printed Scale row. Adds HIPAA-ready BAA, SSO, audit logs, IP allowlisting, private Slack, support hours 8–5. Audit log streams +$128/month.
• Enterprise — contact. Single-tenancy, BYOC, private networking, CMEK (per namespace), 24/7, support + uptime SLA (99.95%). ≥ $4,096/month and a 35% usage premium (public card).
• Usage (all plans) — Storage ≤ $0.33/GB. Writes ≤ $2.00/GB (10 KB minimum per write; batch discount up to 50% at ~3.2 MB). Queries ≤ $1.00/PB of data queried + returned. Namespaces included on the public calculator. Non-filterable attributes are billed at a 50% discount.
• Public calculator example (1,024-d, 1M docs): storage $0.88 + writes $4.71 + queries $1.31 = $7, then the Launch floor lifts the invoice to $16.

Logical bytes (what you ingest), not physical amplification. Confirm checkout.

Limitations / costs

• There is no free hobby cluster. Launch’s floor is $16/month even when the calculator prints $7.
• Cold queries are slower. Public p90 on a 1M-doc cold namespace is about 1.2 s; writes to object storage are not in-memory-fast (p90 ≈ 248 ms for 512 KB upserts).
• A search engine, not an embedding API. For hosted vectors + a notebook see Nomic. For an embedding meter see Voyage AI.

Alternatives

• Qdrant — a filterable vector engine when you want OSS you can run yourself.
• Chroma — a local embedding store when the corpus fits a laptop.
• Nomic — Atlas / embedding tools when the job is a map, not object-storage search.

Official tutorial

turbopuffer’s public channel is @turbopufferdb. Conference talks (including the CTO’s “trillion-scale search” session) are talks, not a first-party getting-started we embed. Use written Introduction plus Pricing.

Website

https://turbopuffer.com

Repository

https://github.com/turbopuffer/turbopuffer-python

Source

Catalog row curated for AI Tools Nav. Cover is the official turbopuffer/turbopuffer-python GitHub file list. Dollars checked on turbopuffer.com/pricing (Launch min $16/mo; storage ≤ $0.33/GB; writes ≤ $2/GB; queries ≤ $1/PB; Enterprise ≥ $4,096). No first-party getting-started embed.

Last verified

2026-10-22 (Asia/Shanghai) editorial check.