There are many vector databases on the market today, and it can be a jungle to navigate. We try our best to publish turbopuffer's limits, performance characteristics, and architecture to make it easy to evaluate. turbopuffer aims to be radically transparent, even if that means you won't choose us yet. If we could do a better job or you're missing a key attribute, let us know.

It is hard to compare fairly. As we attempt to outline fundamental compromises we've made to vector database alternatives to ease your evaluation. Check out limits and performance characteristics.

turbopuffer currently excels at multi-tenancy use-cases where each tenant has its own namespace. Since each tenant is simply a prefix on object storage, this natural sharding scales to billions of tenants with trillions of vectors. Especially if tenants are occasionally dormant, the serverless model is a perfect fit.

turbopuffer may currently not be the best fit for very large namespaces (100M+) that require high QPS and low latency (e.g. marketplaces). They work, but may be slower than our competitors. We're working on improving this. If this prevents your use-case, please reach out.

Pricing for 10M vectors (768 dimensions)$10/month storage + $10 one-time write costAll $100-1000/month
Pricing for 1 QPS (768 dimensions)$10Included
Pricing ModelUsage-based (pay per operation)Time-based (pay to keep your instances running)
Scales to 100,000s of NamespacesYesHave to manually balance between shards/VMs
Elastic ScalabilityYesNone
Scale to ZeroYesNo
Regionsus-central (others on request)Some hosted products let you select a region
Runs on EdgeNoNone
Max Namespace Size~10M vectorsMany undocumented, ~8-40M
Warm P99 Latency for 1M vectors (768 dimensions)75ms, [faster soon][rm]10-200ms (depending on plan)
Cold P99 Latency~350ms10-200ms
AttributesYesYes, generally numbers, bools, strings
Recall90-95%, measure with endpointAll undocumented with no easy way to measure, ~95-98%?
Geospatial SearchNoMost
Built-in embedding creationNoMost
Hybrid SearchSoonSome
Official Client LibrariesPython, more soonAll
Multiple Vectors per DocumentNoSome
Embeddable into your applicationNoFew
Ongoing Recall MeasurementYesNone?
Email us
© 2024 turbopuffer Inc.
Privacy PolicyTerms of service
SOC2 Type 1 certified