Connecting AI to your data · Layer 4
Fine-tuning
Further training an existing model on your own examples to change its behaviour for good: its style, its format or a specific task.
A way to picture itFine-tuning is intensive training that changes a colleague’s reflexes; RAG is handing them the right binder to consult. You do not retrain someone every time a procedure changes.
In practiceStart with prompts and RAG; keep fine-tuning for a stable behaviour that neither of them can deliver.
An insurer wants its claim rejection letters always to follow the same structure, the same tone and the mandatory statements. Despite detailed prompts, the model still regularly strays from them. It is fine-tuned on 2,000 approved, anonymised letters, while the current policy terms are still supplied through RAG. Each new version of the model is run against a set of test cases before going live.
An organisation that fine-tunes a model with its own data becomes responsible for that processing (France · CNIL, 18 July 2024), and a model trained on personal data is not anonymous by default (EU · EDPB, Opinion 28/2024). Learned data cannot be deleted like a database row: minimise or anonymise the examples before training (EU · GDPR, art. 5). Anyone who modifies a general-purpose model using more than a third of the compute of its original training becomes its provider, under the indicative criterion set by the Commission (EU · European Commission, July 2025; EU · AI Act, art. 53). Ordinary fine-tuning stays well below that threshold, but the GDPR applies as soon as one example contains personal data.
“What measurable gain does fine-tuning bring on our own test cases compared with better prompts and RAG, and how would you remove from the model data that someone asks us to erase?”
Poorly. The model retains tendencies rather than precise facts, and can neither cite its source nor keep up with updates. To answer from documents, RAG is the better fit.
Yes, that is one of the advantages of open-weight models: the training data never leave your premises. You do, however, need the right skills and computing infrastructure.
Not in itself. It makes the model more consistent on what it has been shown; reliability is then measured on a set of test cases kept separate from the training examples.
Educational explainer, not legal advice. Up to date as of 29 October 2026.