Cloud-free: it stays in-house
The local AI model runs on your own server. Documents, questions and answers never reach an outside AI provider – so nobody else can learn from them.
Cloud-free AI · a RAG system on your own server
We build an AI assistant (a RAG system) that answers from your manuals, policies, price lists and contracts, and shows the source next to every answer. The model runs on your own server: your documents and questions never go to an AI cloud. A knowledge base for your team, and round-the-clock customer service for your clients.
Eight industries, eight everyday questions – and the answer, with sources.
Asked by: Machine operator, night shiftInternal knowledge base
Machining centre no. 3 is showing error code E-217. What should I do?
Assistant
E-217 signals low coolant pressure Sources: 1. Stop the spindle and check the filter at the back of the machine: if it's clogged, the replacement steps are in point 4 of the maintenance instruction Sources: 2. If the filter is clean, don't restart – call the shift supervisor; in March the same error ended with a pump replacement Sources: 3.
Sources
There's someone to ask even at night: the answer comes from your own manual, with the page number, and the experience of past faults isn't lost when a colleague leaves.
Asked by: Site manager, on siteInternal knowledge base
How thick is the insulation on the attic floor of the Fő utca block? Who approved it?
Assistant
The specification originally required 15 cm Sources: 1, but at the meeting on 6 May it was changed to 20 cm, approved by the client Sources: 2. The material is item 3.12 of the revised cost estimate Sources: 3.
Sources
On site, from a phone, you can see what the current decision is and who made it – no need to call the office, and no building to the old plan.
Asked by: New colleague, second weekInternal knowledge base
How do we record a client's travel expense form, and who has to approve it?
Assistant
According to the firm's procedure, the form goes in the client's folder under “Expenses”, and you record it in journal 2 Sources: 1. Above HUF 100,000 the team lead approves it Sources: 2. A completed example is on slide 3 of the training material Sources: 3.
Sources
New colleagues work independently sooner, and experienced ones don't have to explain the same thing every day.
Asked by: Trainee lawyerInternal knowledge base
Have we ever had a commercial lease where the tenant gave a bank guarantee instead of a deposit?
Assistant
Yes, two: one from 2024 Sources: 1 and one from 2025 Sources: 2. In both, point 9 governs calling the guarantee; in the 2025 one the tenant must top it up within 30 days. The firm's recommended wording is in the contract template Sources: 3.
Sources
Past work can be reused without digging through folders – and only people with access to the matter can see it.
Asked by: Waiter, Friday nightInternal knowledge base
The guest is coeliac and eats lactose-free. What can I recommend from today's menu?
Assistant
From today's menu, the grilled salmon with steamed vegetables and the roast duck leg with red cabbage are gluten- and lactose-free Sources: 1. The mash with the duck is made with butter; steak fries can be asked for instead Sources: 2. Of the desserts, only the sorbet qualifies Sources: 1. Mark the order as an allergen order so the kitchen prepares it separately Sources: 3.
Sources
The waiter gets an exact answer at the table without running to the kitchen – always based on today's menu.
Asked by: Guest, 11 pm, on the websiteCustomer service
Can I bring my dog? And how late can I check out?
Assistant
Pets are welcome in the ground-floor rooms opening onto the garden, for HUF 5,000 extra per night Sources: 1. Late check-out is available until 2 pm, confirmed by reception Sources: 2. You can mention your pet in the comments field of your booking.
Sources
Questions asked at night get an instant answer, and reception doesn't spend the morning answering the same emails.
Asked by: Shopper in the webshopCustomer service
If the shoes don't fit, can I send them back? Do you ship to Austria?
Assistant
Yes, you can return them without giving a reason within 14 days of delivery; the steps are in point 8 of the terms Sources: 1. We ship to Austria in 3–5 working days for HUF 3,990, free over HUF 40,000 Sources: 2. The size chart on the product page helps you choose Sources: 3.
Sources
Fewer repetitive emails, and shoppers don't abandon their basket over an unanswered question.
Asked by: Receptionist at a salonInternal knowledge base
The guest is 18 weeks pregnant. Which treatment can I book for her, and what should we watch out for?
Assistant
According to the salon's treatment guide, pregnancy massage is recommended for expectant mothers from the second trimester, lying on the side Sources: 1. Hot stone treatment and several essential oils are not recommended Sources: 2. Before booking, ask the guest to fill in the pregnancy guest declaration Sources: 3.
Sources
Even a new receptionist answers with confidence, and guests are always informed according to your salon's own rules.
Illustrative examples: the companies, documents and figures are made up. Your system works from your own material and shows its sources the same way.
Not a generic chatbot, but an assistant that knows your company.
The local AI model runs on your own server. Documents, questions and answers never reach an outside AI provider – so nobody else can learn from them.
PDF, Word, Excel, scanned paper (with text recognition), your website and your shared folders. If there's no source for something, it says it doesn't know – and tells you who to ask.
Document, page or clause: one click shows where the statement comes from.
Permissions by group: reception sees something different from management, and the customer-facing chat only works from public material.
As a chat window on your website, and for your team on an internal, password-protected page – on phones too.
When a new price list or policy arrives, you upload it and it answers from that; the old version doesn't get mixed in.
We look at which questions keep coming up, which documents hold the answers, and what size of server it needs.
A first version is built on part of your documents, and your team tries it out.
We measure accuracy on a test set built from your own questions and tune it until it answers correctly, with sources.
We launch it on your own server, connect it to your website or the internal page, and watch what people ask – where an answer is missing, we fill the gap in the documents.
RAG (retrieval-augmented generation) is an AI approach that first searches your documents for the answer to a question, and then writes the reply only from the passages it found, with references. That's what makes it accurate and checkable.
It means the AI doesn't run at a cloud provider abroad, but on your own server, inside your company. We use a local AI model (on-premise): your documents, questions and answers never leave the company or reach a third party.
ChatGPT answers from general knowledge, and whatever you type goes to the cloud provider. A cloud-free RAG system works from your internal rules, prices and manuals, shows the source, and runs on your own server – essentially your company's own ChatGPT, in-house.
We build it to answer only from the sources it finds, and to say so when there's none. Nobody can honestly guarantee this completely – that's why every answer shows its source, and why we measure it with test questions before launch.
Nowhere: documents, questions and answers stay on your own server; no outside AI provider sees them or trains a model on them. Who can access what inside the company is set by group. Because we maintain the system, our access is covered by a data processing agreement.
A suitable machine at your company, yes; an IT person, no. During the assessment we tell you what size of machine it needs – it depends on the amount of documents and how many people ask at the same time – and we help with buying and setting it up. We install the system and maintain it remotely; all you need to do is upload new material when it changes.
In a RAG system, answer quality mostly depends on finding the right passage in your documents – that doesn't need the biggest cloud model. Today's open models that run locally (local LLMs) handle this well; how accurate it is for you, we measure with your own questions before launch.
PDF, Word, Excel, PowerPoint, your website copy, material stored in email or shared folders, and scanned paper too, with text recognition. The better organised the documents, the better the answers – we help with that during the assessment.
Yes: as a chat window on your website, it answers shoppers from your public material (terms, price list, FAQ, house rules), day and night. Where there's no source, it gives your contact details so no question goes unanswered.
After the assessment you get an exact quote in writing, with no obligation. The price depends on the amount and condition of the documents, the number of users, the size of the server, and whether it's for your team or your customers.
Tell us which documents you work with and who asks the questions – you'll get in writing what a cloud-free RAG system could do for you, and what it would cost.