Domain Data

Every specialized field has its own texts, numbers, and images. A hospital collects findings and X-ray images, a law firm collects judgments and contracts, a factory collects measurement data from its machines. Such collections from exactly one specialized field are called domain data. The word domain here simply means subject area or specialized field. The opposite of this is general data: Wikipedia articles, forum posts, news sites — that is, material on all kinds of topics. For programs that learn from examples, this difference matters a great deal.

Why specialized knowledge is not found on the open web

Large language models, that is, programs that continue texts and answer questions, were trained on huge amounts of freely available text. This allows them to do surprisingly much. But the web mainly knows general knowledge. Exactly how a particular turbine reacts at a given pressure is not written in any blog. It is written in the manufacturer’s measurement logs.

This is precisely where the economic value lies. Everyone has access to public texts, including every competitor. Only a single company has its own specialized data. That is why it is now regarded as a kind of raw material. Companies like Siemens, Bosch, or Bayer advertise the fact that they possess decades of measurement data and studies.

A second point is reliability. In medicine or law, a fabricated answer must not happen. If a model draws on verified specialized data, the risk of freely invented statements drops significantly. It does not disappear entirely.

From the filing cabinet to the trained model

There are two main ways to make specialized data usable. In the first, an already finished model is further trained with the specialized material. This is called fine-tuning. The model retains its language ability and additionally learns the terms and patterns of thought of the field.

The second way leaves the model unchanged. Instead, the specialized documents are placed in a searchable repository. When someone asks a question, the system first retrieves the matching documents. Only then does the model formulate the answer, and it may use only these documents. This procedure is called Retrieval Augmented Generation, or RAG for short. It is cheaper, and one can see where an answer comes from.

The most difficult part is usually the preparatory work. Specialized data exists as scanned PDFs, in old databases, or in tables that no one can explain anymore. It must be made readable, sorted, and cleared of errors. Experts estimate that this step often consumes the largest part of the effort.

Domain data in company announcements and products

In business news, you frequently encounter the term in connection with collaborations. An AI provider supplies the technology, an industrial corporation supplies the data. Announcements about such partnerships appear almost weekly. In acquisitions, too, it is often less about software and more about the data holdings of the acquired company.

In products, the principle is present in every assistant that provides information on only a specific topic. A bank’s chatbot knows its rates, not world history. A programming assistant trained on a company’s code suggests solutions in that company’s style.

A common misconception is that more data is automatically better. With specialized data, the opposite tends to be true. A few thousand carefully vetted examples often achieve more than millions of unsorted files. On top of this come legal limits: patient records and personnel data may not simply be fed into training. Data protection also determines which specialized data can be used at all.