ChatGPT on technical manuals? Why a generic LLM is not enough for machine support

In this article we look at what happens when technical manuals are uploaded to a generic LLM, how an agent built on the manufacturer's documentation differs, and when a generic tool is enough instead.

Why machine manufacturers are looking for an AI for technical manuals

Machine manufacturers are looking for an AI for their manuals because service answers the same questions every day, and the answers are already written down somewhere. The problem is finding them quickly.

A technical manual often runs to more than 300 pages, has an index designed for the people who wrote it and uses terms that differ from those of the people who consult it. Every product family has several models, and every model several versions with small but decisive differences: an error code that changes meaning, a replaced component, an updated cleaning procedure.

The result is familiar to anyone who works in after-sales support. Technicians and end customers call service instead of consulting the manual, expert technicians spend the day answering repetitive questions and part of the knowledge stays in the heads of a few people. When one of them leaves the company, that knowledge leaves with them.

Generative artificial intelligence looks like the natural solution: a system that reads the manuals instead of people and answers in natural language. The real question is which kind of AI to use.

What happens if you upload technical manuals to a generic LLM

A generic LLM is a language model trained on all kinds of texts, designed to answer any question. When manuals are uploaded to it, the system usually splits the documents into fragments, searches for those most similar to the question and uses them to compose the answer.

For isolated, well-phrased questions the result is often good. The difficulties emerge in three situations typical of technical support.

The same question applies to different models. A technician asks how to clear error E12. The code exists in three models, with different causes. A system that searches for the most similar text may return the correct procedure, but for the wrong model, and do so in a confident tone.

The information is in a table or an image. Error codes, spare part compatibility and technical specifications are almost always in tables, and procedures are understood by looking at a diagram. If the document is broken into text fragments, tables and images lose their link with the context.

People asking use different words from the manual. An operator talks about a "drawer" where the manual says "collection compartment", a technician uses the internal name of a component. If the system does not know the company's jargon, it does not find the information even when it is there.

On top of this there are two organizational limits. The first is control: a generic LLM answers from any uploaded document, without a review step by whoever is responsible for technical content. The second is context: the model does not know who is writing, which machines they own and what state those machines are in.

The differences between Genuin and a generic LLM on technical manuals

Genuin is the Things5 AI agent that answers technicians, service centers and end customers about a manufacturer's machines. The manufacturer configures it with Genuin Studio: in the Knowledge area it uploads manuals, data sheets and troubleshooting procedures, which become the basis of the answers. The differences compared with a generic LLM stem from how that documentation is organized and from what Genuin knows about the person asking.

1. It starts from the model catalog, not from the documents

Genuin automatically links each document to the machine model or variant it refers to. In the Knowledge area the manufacturer maintains a catalog of its models, with aliases for informal names and the most common typos, which can also be imported from an Excel or CSV file.

When a question is ambiguous, Genuin first asks which model it refers to and shows the options, instead of picking the most similar text. It is the most direct way to avoid the typical error of technical chatbots: the right answer for the wrong product.

2. It reads manuals by chapters, tables and images

Genuin organizes each manual by chapter instead of breaking it into text fragments. Technical tables become usable data and images stay linked to the paragraph they refer to.

In practice, when a technician asks for a procedure they get the complete topic, with the right diagram or photo next to the step that requires it. The manufacturer uploads the PDF and Word files it already has, without rewriting them in a special format.

3. It cites the source of every answer

Every Genuin answer shows the document and section it comes from, with a link to open the source. The technician does not have to take it on trust: they can check in one click, and whoever is responsible for the content can see where a wrong answer comes from.

4. It knows which machines the person asking has

On channels where the user is recognized, Genuin knows which machines that customer or service center owns. In many cases there is no need even to ask for the model: the answer is already tailored to the right product.

If the installed base is connected with Things5 Connect, Genuin also reads the alarms, parameters and history of that specific machine. A generic LLM, by design, only knows the text it is given.

5. It understands the jargon of the company and its customers

In the Knowledge area the manufacturer defines groups of synonyms that link the language of technicians and customers to the official terms of the manual. If an operator uses an internal or colloquial name, Genuin still finds the right information, without the manuals having to be rewritten.

6. It answers in the language of the person asking

Genuin answers in the user's language even when the manual is written in another language. For a manufacturer that sells abroad, this means offering support in several markets without translating all the technical documentation.

7. No content reaches users without approval

In the Knowledge area every piece of content moves from draft to review to published, and only published content is used to answer. Every change to existing content requires a reason, so there is a record of who changed what and why.

Before release the team tests the assistant in the Playground, asking questions as a customer would and checking the sources used for each answer.

8. It improves based on user feedback

Users rate every answer with one click and can leave a comment. Genuin links each piece of feedback to the document that generated the answer, so the team knows which manual needs fixing instead of having to guess. The console also collects all conversations from every channel: useful information for product and marketing too, because it shows what customers really ask.

9. It goes beyond manuals: machine data and business software

A generic LLM answers with text. Genuin can connect, through MCP connectors or APIs, to ERP, CRM, management systems and e-commerce, and with Things5 Connect it can read the status of the machine and act on it within the limits set by the manufacturer. The same conversation can start from an error code and end with the order for the right spare part.

Genuin and a generic LLM compared

The table summarizes the differences for those who have to choose how to automate technical support on product documentation.

When a generic LLM is enough

A generic LLM is enough when the catalog is small, the documentation is stable and the assistant is mainly for the internal team. For example, a manufacturer with a single model and one manual per language can get useful answers by uploading the documents to a generic tool, as long as someone checks the answers before using them.

The picture changes when at least one of these conditions is true:

• the manufacturer has several models or several versions of the same model, with similar manuals;
• the assistant has to answer end customers, dealers or external service centers, so a wrong answer has a cost for the brand;
• the manuals contain tables of error codes, compatibility or diagrams that must be read correctly;
• customers are in several countries and speak languages other than that of the manuals;
• service needs to know which machine the customer has and what state it is in.

In these cases the main problem is not the language model, which today is of good quality in almost every tool. It is how knowledge is organized around that model: catalog, versions, approvals, customer context.

Seven questions for choosing an AI assistant for technical support

Whether you are evaluating Genuin or another solution, these questions help you understand whether a tool is suited to a machine manufacturer's documentation. It is worth asking them during the demo, using a real manual and a question that service receives often.

1. How does it distinguish models and versions? Ask what happens if the same question applies to two different models.
2. What does it do with tables and images? Try a question whose answer is in a table of error codes.
3. Does it cite the source? Check that every answer points to the source document and section.
4. Does it know who is asking? Ask whether it can link the user to the machines they own.
5. Who approves the content? Check whether there is a review step before content is used with customers.
6. How is quality measured? Ask how feedback is collected and whether it is linked to the documents.
7. Can it go beyond text? Check whether it can read machine data and connect to ERP, CRM or e-commerce.

To learn more about reducing repetitive requests to service, also read how to reduce repetitive calls to technical service.

Frequently asked questions

The difference lies in how knowledge is organized

Today language models write well almost everywhere. For a machine manufacturer, the quality of an AI assistant depends on what the model knows about the product: which machine, which version, which customer, which approved document. A generic LLM starts from the uploaded texts. Genuin starts from the model catalog and the installed base, and uses the manuals within this structure.