Fine-tuning

Fine-tuning is used to change a model's behaviour: teaching a particular tone, a specific output format, or a domain-specific phrasing pattern. It is a tool designed for shaping behaviour, not for adding knowledge.

This distinction is often blurred. The sentence "let's teach the model our company documents" usually points at fine-tuning, but what it describes is RAG. If the goal is to add knowledge, the right tool is a retrieval-based approach.

Fine-tuning is not a one-off job. Every time a document changes, the training set must be refreshed and the process repeated. Since procedures in SMEs change several times a month, that cycle stops being sustainable.

With RAG the same change ends with re-indexing a single document and takes minutes. The cost difference is not only in the training fee but in the maintenance effort — and at SME scale the second is decisive.

There is no way to show the source of an answer from a fine-tuned model; the knowledge is dissolved into the weights. In an audit you cannot answer "which procedure is this answer based on".

Incorrect information is also hard to withdraw: if a wrong document entered the training set, removing it requires retraining. With RAG the same operation ends by taking the document out of scope. In regulated processes that is a decisive difference.

A useful way to evaluate fine-tuning is to follow one real question from the source document to the final answer. Record which file was selected, what context reached the model, and what a reviewer would need to verify. This turns a definition into an operational check and makes the result comparable across teams.

The same check should include ownership and change management. Decide who updates the relevant documents, how an outdated result is reported, and which access boundary applies. KobiGPT can provide the assistant and the cited document context, but the organisation still owns the source material, permissions, and the decision made from the answer.