Prompting vs RAG vs Fine-Tuning

One rule sorts the three main ways to improve a model's answers: prompting changes the instructions, RAG changes the knowledge, and fine-tuning changes the behavior. Fine-tuning looks like the exception. It trains the model's learned numbers further, and those numbers do store facts, so many engineers expect fine-tuning to teach a model new facts as well. It can store a few, but those facts are hard to check and slow to update.

A store's support assistant told a customer her new blender had a one-year warranty, but the store had extended it to two years