An AI system only becomes useful for a company once it knows its own procedures and works with real business data.Off-the-shelf models do not deliver this: they are trained on generally available texts and know nothing about your product logic, your proposal templates or your internal approval workflows.So they must be supplied with context.Two approaches are up for discussion for this step – Retrieval-Augmented Generation, or RAG for short, and fine-tuning.Which one is the right fit depends on the use case.

What RAG does

With RAG, the language model remains unchanged.A search layer is placed in front of it: for every request, matching excerpts are searched from your own holdings – manuals, wiki pages, ticket histories, product data – and passed to the model together with the question.The answer is therefore produced on the basis of material that is provided only at the moment of the request.

The practical advantage lies in currency.If a policy changes, only the document in the holdings needs to be replaced; the model itself remains untouched.It is equally possible to trace which sources a statement comes from – a point that carries considerable weight in regulated environments and when providing information to customers.Access rights can also be mapped at this point: anyone who is not allowed to see a document does not receive it as context either.

The effort instead shifts to data preparation.Documents must be indexed, sensibly broken down and made searchable.If the holdings are unstructured, outdated or contain duplicates, even the best search layer will deliver weak hits – and the model will formulate a fluent but incorrect answer from them.

What fine-tuning does

With fine-tuning, the model itself is retrained, using example material from the company.The knowledge thus moves into the model's weights, not into an upstream database.The approach plays to its strengths where the focus is less on facts than on behaviour: a particular tone, a fixed answer format, domain-specific language quirks or recurring classification tasks in which the same kind of input should always produce the same kind of output.

The price for this is inertia.New or corrected knowledge does not enter the system via a file upload but requires a new training run with appropriately prepared examples.For content that changes frequently, this is the wrong lever.In addition, with retrained models it is considerably harder to prove where a specific statement comes from.

The decision along the use case

Instead of a fundamental preference, a look at the characteristics of the case helps:

  • How quickly does the knowledge become outdated? Daily prices, stock levels or policies argue for RAG.
  • Is it about facts or about form? Factual information argues for RAG, consistent format and style for fine-tuning.
  • Must the origin of an answer be verifiable? Then a source citation from the search step is hard to replace.
  • How homogeneous are the tasks? A narrowly defined, often repeated process is more worth the training effort than an open question-and-answer scenario.
  • What data basis exists at all? Fine-tuning needs clean example pairs, RAG needs well-maintained documents.Both are work, but different work.

In practice, the approaches are not mutually exclusive.A model can be adapted to format and domain language and still draw its substantive content from a search layer.Anyone who wants to start small usually begins with RAG: the path is more reversible, the results can be verified sooner, and the data preparation pays off even if fine-tuning is added later.

What this means for operations

Both variants create ongoing effort, just in different places.RAG systems need functioning maintenance of the knowledge base: who is responsible for ensuring that an outdated document actually disappears from the index?Fine-tuning approaches need a defined process for new training runs and a way to detect quality degradation before it becomes visible in production.

In both cases, this includes an evaluation that goes beyond random samples – a fixed set of typical questions with known correct answers, against which every change is checked.Without this benchmark, it is impossible to determine whether an adjustment has improved or worsened the system.

The CMS connection

For content management systems, the point is more obvious than it initially sounds.A CMS is already a structured, access-controlled, versioned knowledge base – exactly what a search layer needs.Editorial assistance, internal search or chat information based on your own content can thus be connected without content being transferred into a model and permanently frozen there.Editorial sovereignty remains where it belongs: in the editorial system.

For web and software projects, this means: the question of RAG or fine-tuning is not purely a model question but a question of data storage and responsibilities.Anyone who already structures, versions and applies rights to their content cleanly has already done the larger part of the preparatory work.

And those who make the architecture decision early avoid a prototype later being convertible into an operable service only with great effort.

Sources