Fine-tuning continues the training of a pre-trained model on a specific dataset, to adapt it to a task, a style or a domain. It modifies the model's weights. It requires quality examples and compute cost, and does not replace a RAG for injecting knowledge that changes.