Private AI for manufacturing: manuals, procedures and maintenance logs, answered in your language

The problem

In a manufacturing company, knowledge sits in three piles. The first is the machine manufacturers' manuals, often in German or English, several hundred pages per machine. The second is your own procedures and work instructions from the quality system. The third is the maintenance logs that record what broke on which line and how it was fixed.

When a line stops at two in the morning, the technician is not going to page through a 400-page German PDF. They call a colleague who has seen the fault before, or they wait until morning. Every hour of downtime has a cost you know better than we do.

The problem grows as experienced technicians retire. Their knowledge of how each machine behaves is often only partly written down, in short log entries nobody reads end to end. Private AI does not replace those people, but it makes their records available to everyone at the right question, including a new colleague on the night shift.

General-purpose chat assistants help little here. They do not know your machines, your procedures or your fault history, and supplier documentation is often confidential under contract, so you may not upload it to a foreign service. In some companies the production network has no route to the internet at all.

What the private AI does

We install a language model on a server inside the company and connect it to your technical documentation through RAG search. A technician, process engineer or operator asks in Slovenian. The system finds the relevant paragraphs in manuals, procedures and logs, even when they are in German or English, and answers in Slovenian, citing the original source. English-speaking teams can ask and get answers in English the same way.

A sample question and answer (made-up documentation, shown for format):

Question: The press on line 2 shows hydraulic pressure fault E-214. What do we do?

Answer: The manufacturer's manual prescribes three steps for fault E-214: check the oil level in the tank, check the filter on the pressure side, and only then the pressure sensor. The maintenance log records the same fault on line 2 twice; both times the cause was a clogged filter, and the replacement took about 40 minutes.

Source: Manuals/Press_HP400_DE.pdf, p. 212 (Fehlercode E-214); Maintenance/Line2_2025.xlsx, rows 88 and 141

Other typical pilot questions:

• What tightening torque does the manual specify for the bolts on the clamping plate?
• Which procedure applies to receiving material from a new supplier?
• When was the bearing on this machine last replaced, and who did the work?
• Summarise the safety requirements from the manual for cleaning the cutting unit.

Data sources

For the pilot we pick one line, one machine type or one set of procedures, so quality can be measured precisely. The most common sources:

• Manufacturers' manuals in German, English or Slovenian, including scans, with text recognition.
• Procedures and work instructions from the quality system, tagged with version and status.
• Maintenance logs from Excel, a maintenance system, or the forms technicians fill in.
• Supplier documentation: data sheets, certificates, material specifications.

The system recognises drawings and schematics as documents and cites them as sources, but it does not interpret what a drawing shows. We are careful here because a misread dimension could do more harm than good.

Architecture

The whole chain runs inside your network, without an internet connection if needed:

1. Indexing: documents are split into paragraphs, page and chapter numbers are kept, and the document version is recorded.
2. Multilingual retrieval: a question in Slovenian finds a paragraph in German or English, because retrieval works on meaning rather than on matching words.
3. Model: an open language model (we choose between the Llama, Mistral, Qwen and Gemma families depending on Slovenian-language quality and hardware) builds the answer from the retrieved paragraphs only.
4. Interface: a web chat on a computer or tablet on the shop floor; optionally an OpenAI-compatible API to connect the maintenance system.
5. Oversight: the question log shows what people search for most, which is a good hint about which manuals need rewriting.

The company buys the hardware; we size it and do not resell it. The options, from an existing server to a dedicated GPU server, are on the private AI page.

Security and regulation

Supplier technical documentation is often confidential under contract, and your procedures and fault history are trade secrets. A private deployment means none of these documents goes to an outside model provider.

For companies with a certified quality management system it matters that an answer points to the current version of a procedure. That is why the index stores each document's status and searches current versions only by default.

For safety-relevant steps (lockout, isolating energy, work at height) the system always shows the original paragraph from the manual, not only a summary. How the work is carried out remains the decision of a trained worker and their supervisor.

Timeline and price

• Pilot (from 2,900 €, 2–3 weeks): one line or one machine type, a local deployment and an evaluation on 30 to 50 real questions from technicians and process engineers.
• Production "Company knowledge" (6,000–15,000 €): all lines, connection to the documentation and maintenance systems, permissions by department, monitoring and onboarding.
• Support (290–790 € per month): adding new machines and manuals, updates and quality monitoring.

The pilot runs on one line, with the people who know it. In the first week we collect the manuals and logs and, with the maintenance manager, put together 30 to 50 questions that actually came up in the past year, along with the correct answers. In the second week we install the system on the server and test it on those questions, checking safety-relevant steps and answers drawn from German or English documentation with particular care. In the third week the technicians on the line use it for a few days, and then you get a report with answer accuracy, a list of errors and an assessment of which line should come next.

Prices are final; the provider is not VAT-registered. What each tier covers is on the private AI for business page. For policies, HR questions and onboarding in the same company, see the internal knowledge base.

What we do not promise

• Reading technical drawings. The system finds and cites drawings; it does not interpret them.
• A replacement for a trained technician. The system shortens the search; diagnosis and repair are done by a person.
• A perfect translation. An answer in Slovenian is a summary of the original. For safety-relevant instructions the original text applies, and the system always shows it.
• The quality of the largest cloud models on every task. Open models on your own server are very good at searching documentation and weaker than the largest cloud models at demanding technical reasoning. We show this in the pilot.