Perimeter arrives as a single appliance. It holds the models, the index and your documents. Every question is asked and answered inside your walls — this page shows how.
Live in 30 days · hardware bought in the EU · model updates included in support
The appliance is standard hardware, bought in the EU under your name. We size the class to your corpus and your users; the default fits most mid-market teams.
A small unit for one team and a focused corpus. Same software, same zero external calls — less machine.
A GX10-class box with large unified memory: it holds a serious open-weight model and your whole index in one unit, in your server room.
Workstation-grade GPU for more users and faster reading. The class the pilot's load test points to when the default is not enough.
An honest note on speed. The GX10's unified memory holds big models, but it reads long documents more slowly than a datacenter GPU would. For a team asking questions and matching lines, it is quick enough — and the pilot's load test proves it on your real usage. When it isn't enough, the heavier class exists for exactly that.
When you load a document, the box stores two things: the file itself, and what every passage means. Ask a question and it first finds the passages that mean the same thing — even when the words differ — then writes the answer from those passages alone. It does not answer from memory. It answers from your documents.
Catalogues, contracts, standards, matters — copied onto the box locally, never uploaded anywhere.
Every passage is stored by what it means, not just the words it uses. "Steel pipe 57×3.5" and its other spellings land in the same place.
The model writes only from the retrieved passages — and hands you the answer with its source attached.
Every answer carries the document and the place it was taken from. You do not have to trust the machine; you open the source and read the line yourself.
And when your corpus does not contain the answer, it says so — nothing is invented to fill the gap.
Open-weight models improve every few months. Perimeter is built so the model can be replaced without touching your corpus or your index. Today the slot holds gpt-oss, Qwen3.5 or Mistral Small 4 — whichever measures best on your documents. When a stronger model ships, we test it on your reference sample and swap it in. Model updates are included in annual support.
The model's weights sit on your box, with public provenance. Nothing about it is a black box you have to take on faith.
Every model in the slot runs entirely on the appliance. No API, no callbacks, no phone-home.
We pick the model that scores best on your reference sample — not the one with the loudest launch.
No live prices, no news, no case-law updates. The box knows the corpus you loaded, and nothing else arrives on its own.
A machine that fetches nothing can leak nothing. You load what it knows — so you always know what it knows.
How many people one box serves depends on your documents and how your team uses them. A number on a website would be a guess.
The pilot includes a load test with your real number of users on your real corpus. The box is sized on evidence, not a spec sheet.
Perimeter joins your rack like any other unit: power, and one cable to your LAN. It needs no route past your firewall — on day one or on any day after.
Twenty minutes, no slides. Bring a contract or a catalogue page — we'll load it and show the answer with its source, live.