Start on shared infrastructure because it is fast and cheap. Move to a single-tenant instance when the workload or the auditor says so. Or run the whole thing inside your own perimeter. It is the same platform at every rung — what changes is who holds the keys.
Every rung runs the same platform, the same retrieval and the same governance. What differs is whose infrastructure it sits on, who operates it, and - on Shared - which plan you are on.
Your own tenant on our multi-tenant platform, with per-tenant corpus isolation. It is running today: sign up, load documents, and point a tool at them the same afternoon.
From EUR 39 per month.
See plans and startA single-tenant instance on hardware reserved for you, operated by us in the EU or your region. Nothing of yours shares a database, a vector store, or a model route with another customer.
Priced to the hardware you actually need — scoped with us.
Scope a dedicated instanceThe platform installed inside your own perimeter — your Azure or AWS subscription, your datacenter, your rack. Licensed to you and operated by your team, by us under contract, or by both.
Licence plus services, on a scoped contract.
Talk about on-premThe reason to start small is that starting small costs you nothing later. Changing rung is a migration we perform, not a rebuild you pay for twice.
Documents, chunks and embeddings are migrated as data. You are not re-uploading a library or re-indexing from scratch.
The same tool surface and the same tokens, repointed at the new address. Whatever you built against the platform keeps calling it.
Retrieval settings, prompts, and any model work done for you are configuration and artifacts — they are portable between rungs by design.
Migration between rungs is an engagement we run for you, not a self-serve button. That is deliberate: it is the step where a mistake costs you your corpus.
We have a stack we prefer, and we will happily argue for it. What we will not do is make it the only thing that works. Every layer a customer could be trapped by is one we deliberately keep swappable.
On Shared and Dedicated, on servers we operate in the EU or your region. On Sovereign, wherever you put it — that is the point of the rung. We will tell you exactly which services hold which data before you sign anything.
Yes, on every rung. Model calls route through a single model gateway, so you can use open-weight models reached through it, bring your own provider key, or mix the two. Swapping a model does not mean rebuilding your corpus.
You keep your documents and your extracted text and embeddings — we export them in open formats. The retrieval and inference layer is built on open-source components, so there is no proprietary data format standing between you and the exit.
That is the intended path, and most sensible for a first project. Prove the use case on Shared where setup is an afternoon, then move once the workload and the compliance requirement are real.
Only what a given piece of work requires, and only for as long as it requires it. On Sovereign we can hand over entirely and stay out, which is how several deployments are already run.
Most of this decision is made by one sentence about where the data is allowed to sit. Bring that sentence and a real workload, and we can scope the rest in a single conversation.