Turbopuffer

Turbopuffer is a purpose-built search database that turns the typical cost model of vector search on its head. Instead of paying for always-on compute nodes, you pay for object storage and a cache. This design gives you dramatic cost savings—the vendor claims 10x cheaper than alternatives—but it's not without tradeoffs. The most visible tradeoff is higher write latency (p90=248ms for 512KB upserts) because writes go directly to object storage. For workloads with high write throughput and less strict latency requirements, this is usually fine. More serious is cold-query latency: when data isn't cached or pinned, p90 can be over 1 second. If your application requires sub-100ms responses on every query, you'll need to pin namespaces (limited to 256 by default) or consider a different tool. Where turbopuffer shines is the retrieval stage of AI pipelines. The team explicitly focuses on first-stage retrieval: narrowing millions of documents down to tens or hundreds. The hybrid search (vector + BM25) works well for this, and the recent additions like namespace branching and sharding make it easy to test changes and scale beyond 128B docs. The native embedding beta is interesting but not yet GA, so you may still need an external embedding model. Operationally, turbopuffer is API-first and multi-tenant by default. You create namespaces and interact via a REST API. Copy-on-write branching lets you clone a namespace instantly—great for testing ranking changes without duplicating storage. The dashboard is getting better but is still minimalist; you'll mostly live in the API and docs. For teams with modest scale (a few million vectors), the minimum commitment of $16/month might be more than, say, a free tier of some alternatives, but for anything substantial, the cost advantage is real. Just be sure to understand that you can't run it on-premise (it depends on S3) and that you get no SQL or ACID transactions—it's a search engine, not a full database.