When to retrain and when not to

Three ways to adapt AI to a task, and they are constantly confused. Prompting - change the instructions, the fastest and cheapest, try first. RAG (search through your documents) - when you need FRESH or private KNOWLEDGE: you pull relevant pieces into context. Fine-tuning (additional training) - changes the STYLE, format and behavior of the model on many examples, but does NOT add knowledge reliably (the model does not remember facts that way). The classic mistake: “let’s train further using our documentation so that you know it” - almost always this is a task for RAG, not for further training. Rule: knowledge is RAG, style/format/behavior is fine-tune, and always start with a prompt.