Two ways to make AI useful
A general-purpose large language model (opens in a new tab) knows nothing about your procedures, customers or vocabulary. Two techniques close the gap: RAG, which hands it the right documents when it answers, and fine-tuning, which retrains it on selected examples.
They are not rivals. In most projects we start with a RAG assistant and consider fine-tuning when style or accuracy is no longer enough.
RAG: answers grounded in your documents
RAG (Retrieval-Augmented Generation) was described in 2020 in a research paper from Meta AI (opens in a new tab). Your documents are split and indexed; for each question, the relevant passages are retrieved and passed to the model with the question.
Its strengths: answers cite their sources, documents can be updated without retraining, and access rights can be respected. It is ideal for internal procedures, contracts and knowledge bases.
Fine-tuning: a way of working
Fine-tuning continues a model’s training on your own examples. Methods such as LoRA (opens in a new tab) make it affordable by changing only a small part of the model.
It helps when the model must adopt a precise style, follow a strict output format or master very specialised vocabulary. It requires a high-quality dataset, built and checked with your experts.
How to choose
- Your information changes often: choose RAG.
- You need to cite sources: choose RAG.
- The model must write in a very precise format or style: consider fine-tuning.
- You want your own model, hosted in-house: combine fine-tuning with self-hosted AI.
What about privacy?
Either way, your data stays under your control if the model is hosted by you or a European provider. The GDPR (opens in a new tab) applies as soon as personal data is involved, in training and in use. We cover this in our article on AI and the GDPR.