AI for the files you can't put in the cloud.
One machine, in your office. It reads everything you hold — decades of it — and answers with the page it came from. It never touches the internet. No exceptions.
Book a discovery visitFor law firms, investigators, accountants, finance and newsrooms — any organisation whose documents can't leave the building.
No cloud accounts·No per-user licences·Nothing leaves your network·One machine, yours outright
What is RAG
Retrieval-augmented generation: retrieve the right pages, then generate the answer from them — never from guesswork.
We build the machine
A quiet tower with a serious AI processor inside, built and tested for your firm. You own it outright — the hardware and everything on it.
It reads your documents
Case folders, contracts, ledgers, email archives, scanned paper. Every page indexed once, in your building.
Your staff just ask
Every computer on your network is a terminal to the machine. They type the question they'd have asked a colleague; the answer comes back with the page it came from.
It works at three depths, and we set yours to match the work:
Look it up
Finds the passage, answers in a sentence, cites the page. It searches by meaning as well as by keyword, so it lands on the right paragraph even when the wording differs. The everyday workhorse.
Ask itWhat do the case files say about this — and where exactly?
Reason it through
A conclusion and the working behind it, every step resting on a cited passage — so you can check the reasoning, not just the answer. It also marks its own homework: if the pages it found don't really answer the question, it goes back and searches again.
Ask itCompare these five documents — where do they differ, and does it matter?
Research it
Give it a big question and it plans its own research — searches the archive, reads what it finds, follows the trail, and comes back with a structured report and a citation on every finding. It will also map how people, companies and documents connect. Everyday questions stay instant.
Ask itWho is connected to whom across this case material — and on what evidence?
Most firms live at Levels 1 and 2 and reach for Level 3 on the hard cases.
Why not just use ChatGPT?
For this material, you can't:
Confidential by law
These documents are privileged, evidential, or covered by data-protection rules. Material like this belongs air-gapped, inside your own walls — sending it to someone else's cloud isn't a preference call, it's a breach.
It hasn't read your files
You can't paste an archive into a chat window. The knowledge that matters to your firm is in your building, not on the internet.
No proof
A confident answer with no source is worthless in court, in disclosure, or in front of an auditor.
Do you need private AI?
Private AI is worth it when what goes in is confidential.
It can be overkill when there's no secret to protect.
The questions count too — a cloud provider sees every query your staff type. This machine tells no one.
The material is the secret — you need private AI
- Summarising a client bundle
- Questioning the contents of a seized phone
- Transcribing a recorded interview
- Extracting figures from audit files
- Drafting from your own precedents
- AI help on your own source code
- Research where the questions themselves reveal strategy
Nothing secret goes in — the cloud is fine
- Writing a marketing email
- General knowledge questions
- Drafting a job advert
- Making an illustration for a slide deck
- Polishing the wording of a public web page
Part of the discovery visit is sorting your workflows into these two columns — including telling you where you don't need us.
How it works
Your documents, the machine, and your staff — all on your own network.
Your documents
Document system, shared drives, email archives, scanned paper. They stay where they are.
The RAGent machine
Holds the index of every document, and the AI that reads and answers.
Your staff
Every computer on your network is a terminal to it. Nothing to install.
- A member of staff types a question.
- The system searches the index and pulls the relevant pages.
- The AI writes an answer using only those pages.
- The answer arrives with citations — click one, see the source page.
Answering questions is one job of many. The same machine will:
- Summarise a bundle
- Extract dates, sums & names
- Compare documents
- Draft from your precedents
- Propose redactions
- Map who's connected
- Answer from your code
- Act in your other systems — tool calling
☁ The cloud: not involved. Unplug the internet and it still works.
The grounding rule — the AI may only answer from the pages it just pulled out of your documents. If the answer isn't in there, it says so instead of inventing one. And every claim carries a citation you can check.
The AI itself is free — and yours
Nobody rents you the brain
Meta, DeepSeek, NVIDIA and Mistral publish models anyone can download and run for nothing — and they are genuinely good.
No per-message charge. No meter.
The model never sees your files
It runs on your hardware, so there's no account, no upload, and no terms of service sitting between you and your client files.
No licence to expire, and no service anyone can switch off underneath you.
It gets better, not older
New models land every few months, and we swap them in on hardware you already own. Its knowledge isn't frozen either — point it at a research library or subscription you licence and that becomes searchable too.
Want to run your own models alongside ours? It's your machine.
Use cases
Police & investigations
Thirty thousand pages off a seized phone, and a deadline. Cloud AI is legally impossible here — the material cannot leave the force's network. This reads all of it overnight and answers with citations that survive disclosure.
Who is connected to whom across this case material — and on what page?
Law Firms
Twenty years of matter files, half of it scanned paper. Fee earners stop hunting through folders and start asking the archive directly — and every answer links to the page, so it can be checked before it's relied on.
Summarise this bundle in two pages, and flag anything inconsistent with our earlier position.
Accountancy & audit
Audit files, ledgers and correspondence in one place. Pull the numbers and dates out of a year's paperwork without losing a junior to it for a week.
List every date, sum and name mentioned across these audit files.
Finance & compliance
Client files and policy you're expected to know cold. Answers come with sources, and an audit log records who asked what, when, and which documents were used.
What does our policy say about this, and when did it last change?
Software development
Banks and regulated firms build software too — and their source code is as confidential as their client files. The machine reads your repositories and internal docs, so developers get AI help without a line of code going to a cloud assistant. They can work from the terminal, too — it's a standard API on your network.
Where is this calculated in our code, and what does it depend on?
Journalism
Leaked archives and long investigations, where the material can't touch a third-party service and a source must never be exposed.
What does the archive say about this person, and where does each claim come from?
Any document-heavy team
If your people spend hours hunting for what the firm already knows, that's the job this takes off them. We'll find the ones worth automating when we visit.
Draft a reply to this letter, consistent with our previous position.
What's actually in the box
Your IT people will ask. It's a workstation-class machine, not a rack of servers.
- AI processorA professional NVIDIA card — how big it needs to be is the main thing we work out when we visit.
- MemoryError-correcting, so a long job can't quietly corrupt itself.
- StorageEnterprise drives for your index and documents, sized to your archive with room to grow.
- SoftwareLinux, open-source models, the search index and a web interface — all served locally.
- Where it livesA 13-amp socket and a network port. Comms cupboard, server room, or under a desk. No rack, no special cooling.
- On your networkSigns in against your existing directory, and respects the matter walls you already have. Every staff computer is a terminal to it — nothing to install.
- Does it phone home?No. No telemetry, no licence checks, nothing calling out. Air-gap it if you want to.
How is this different?
Cloud AI tools are good at what they do. They just can't do this.
| ChatGPT / Claude | Microsoft Copilot | RAGent | |
|---|---|---|---|
| Where your documents go | Their cloud (US) | Microsoft's cloud | Nowhere. They stay in your building. |
| Has it read your archive? | No | Partially — what lives in Microsoft 365 | Yes — all of it, including scanned paper |
| Cites the exact page? | No | Links to files, loosely | Yes — every claim, click through to the page |
| How it's priced | Per user, monthly, forever | Per user, monthly, forever | Buy the machine once. Whole firm uses it. Fixed monthly support. |
| Works offline / air-gapped | No | No | Yes |
| Best at | General knowledge, drafting from scratch | Email and Office documents | Your confidential archive, with proof |
If you want help writing marketing emails, use Copilot — it's good at that, and we're not competing with it. RAGent is for the documents that can never go where Copilot lives.
What it won't do
Straight answers, up front.
It won't replace your judgement
It finds, drafts, summarises and cites — fast. Your professionals still decide what's right and put their name to it. It makes the expert quicker; it doesn't pretend to be one.
It won't ask you to take its word
No AI is perfect. Every claim carries a citation you can click. You never have to trust the machine — only the page it shows you.
It won't lock you in
You own the machine and everything on it, and the software is open source. If we vanished tomorrow, it keeps working.
Products
Three machines, built to order — same software, sized to how many people ask at once and how hard they make it think.
Desk
5–20 staff · one team
A small practice, or one team inside a bigger firm. Sits on a desk and answers a few people at a time.
Practice
20–100 staff · firm-wide
The whole office, all day. Deeper reasoning, a bigger archive, and room for everyone to ask at once.
Vault
100+ staff · heavy caseloads
Litigation-scale document sets and deep research all day. Runs the largest open models we trust, NVIDIA's Nemotron among them.
Pricing is fixed, and in writing. It starts with a free half-day discovery visit: we go through your document work, size the right machine, and send you one quote covering hardware, setup, training and support. No per-user fees. Five users or five hundred — you bought the machine.
From first visit to first answer
A delivery, not an IT programme.
Discovery visit half a day · free
We sit with you, work through how your documents actually move, and size the machine.
Fixed quote within days
One document: hardware, setup, training, support. In writing.
Build & test 2–3 weeks
We assemble it, install the software, and run it hard before it ever sees your building.
Install & index on site
Power socket, network port, done — no screen, no keyboard. We point it at your document stores and it indexes overnight, or over a few days for a big archive.
Training & handover
We train your staff in the room, on your real documents. Then it's yours.
Book a discovery visit
Half a day at your offices, free — see the system answer on your real files, then a fixed quote in writing.