Definition
Partial retraining of an AI model on specific data to specialize it for a domain, tone or task.
Fine-tuning consists of continuing the training of an existing model on a targeted corpus, internal case law, customer support history, technical documentation, to improve its performance on that specific domain.
For most business uses, RAG is enough and remains simpler to maintain (knowledge is updated by re-indexing documents, without retraining). Fine-tuning keeps its value for specializing a local open-source model on a domain vocabulary or a highly constrained output format.
A law firm retrains a model on 2,000 of its own letters so that it adopts the firm's style and vocabulary. The model then writes first drafts that read like the firm's own.
When you need a very specific and stable tone, format or vocabulary. To give the AI access to your documents, RAG is almost always simpler, cheaper and easier to keep up to date.
Do I need to fine-tune a model for it to know my company?
Rarely. In most cases, connecting the model to your documents (RAG) is enough and avoids retraining it at every change. Fine-tuning is justified for style or for highly repetitive tasks.
All AI models, deployed on your infrastructure, with your documents connected.