Documentation

Five minutes from empty account to first query.

There is one API surface: a namespace you write to and query. Everything else — embedding, reranking, consistency, filtering — is a parameter on those two calls.

{{ fileName }} 
{{ code }}
Concepts

Six ideas and you know the system.

namespaces

The unit of everything

Isolation, billing, consistency and caching are all per namespace. Create one per tenant — a hundred million of them is a normal deployment, and idle ones cost nothing.

 consistency

Strong, bounded, eventual

Per query, not per cluster. Strong reads see your last ack; bounded reads state their own upper limit on staleness; eventual reads skip the check entirely.

 durability

Standard and realtime writes

Standard writes are visible in about a second at $2/GB. Realtime writes commit synchronously across two zones in roughly 25 ms at $3/GB. Choose per call.

 hybrid search

Vector, BM25, filters

One index format holds all three. Fuse dense and lexical results with reciprocal rank fusion server-side, or take both result sets and fuse them yourself.

 branching

Copy-on-write namespaces

Segments are immutable, so a branch is a new manifest pointing at existing objects. Instant, free until you diverge, and safe to hand to a staging environment.

 pipeline

Embed and rerank in-request

Send text instead of vectors and we embed it. Ask for reranking and a cross-encoder reorders candidates before the response leaves our network. Both are optional.
API reference

The whole surface area.

OpenAPI spec →
{{ e.method }} {{ e.path }} {{ e.desc }}
Guides

Built for the workloads people actually run.

Agent memory that survives the session Write every turn with realtime durability, read it back strongly consistent on the next turn. Includes the TTL and compaction settings we recommend. → One namespace per tenant, at scale Provisioning, quota and noisy-neighbour patterns for B2B SaaS with tens of millions of tenants, most of them idle most of the time. → Migrating off Pinecone or Turbopuffer A dual-write window, a backfill job and a recall comparison harness — so you can cut over on evidence rather than optimism. → Keeping recall under heavy filters Why centroid indexes lose recall on selective filters, what our filter-aware probe does about it, and when to force the exact fallback. →

Run it locally first.

The single-node engine is Apache-2.0 and speaks the same API and the same segment format. Develop against a binary on your laptop, deploy against the managed service.

Install the OSS engine Create an account