Nexuldra
// ai
Fine-tuning
Fine-tuning is continuing a model's training on your own examples — hundreds to thousands of input-output pairs — so its weights shift and it reliably produces that style, format or behaviour. It changes how the model acts; it does not add facts. For feeding it your documents, RAG is the right tool; fine-tuning serves style, format and tone.
Fine-tuning matters because prompting alone sometimes cannot lock in a behaviour: a consistent output format across thousands of calls, a house style, a classification scheme the model keeps drifting from. A tuned model also shortens prompts — the behaviour is baked into the weights — which cuts latency and cost per call. The economics only work at volume: preparing a clean dataset, running training and evaluating the result is an engineering project, not a configuration toggle.
Concretely: a support desk that wants every reply in its exact format — greeting, diagnosis, steps, signature — can fine-tune on a few hundred past good replies and stop babysitting the prompt. Prices have fallen to roughly tens of euros per training run for modest datasets on hosted platforms. But when the underlying facts change weekly, fine-tuning is the wrong tool: retraining to update a price list is absurd, while a RAG index updates in minutes.
Frequently asked questions
- Fine-tuning or RAG — which do we need?
- For answering from your documents: RAG, almost always — it is updatable in minutes, cites sources and cannot forget. Fine-tuning when you need a stable format, tone or classification behaviour across thousands of calls. Many systems end up using both: retrieval for facts, tuning for form.
- How much data does fine-tuning need?
- Hundreds of high-quality examples often beat thousands of mediocre ones. The dataset must be consistent: if your examples contradict each other, the model learns the contradiction. Budget serious time for cleaning — data quality, not model choice, decides the result.