Fine-tune, prompt, or RAG?

Why this matters

The most expensive mistake in this whole field is fine-tuning when you didn't need to. Fine-tuning costs you a dataset, a training run, an eval harness, and a deployment — weeks of work — and people reach for it reflexively when a better prompt or a retrieval step would have solved the problem in an afternoon. The three tools solve different problems: prompting changes instructions, RAG injects fresh or private knowledge at query time, and fine-tuning bakes behavior into the weights. Knowing which lever to pull before you start is the single highest-leverage skill in this course, because it decides whether you spend an afternoon or a month.

Demo

The demo is a decision function, not a model: given what you're actually trying to change (style, format, latency, or new facts), it tells you which tool fits — so you stop reaching for fine-tuning by default.

Try it yourself

1. Run choose_approach on a real problem you have right now and write down which tool it points at.
2. Flip facts_change_often from True to False and watch the answer move from RAG toward fine-tuning.
3. Find a case where the right answer is 'all three' (RAG for facts + a fine-tune for format + a tight prompt) and justify each layer.
4. Identify one problem you previously assumed needed fine-tuning that a better prompt would have solved.

Prompt your AI

Use these three in order. Each builds on the one before.

1. Basics & terminology

In one paragraph, explain the difference between prompting, RAG, and fine-tuning, like I'm new to it.

2. Why it works (the mechanism)

Walk me through how I'd decide between fine-tuning, RAG, and prompt engineering for a given task, step by step.

3. Advanced — application & what's next

Given a customer-support bot that must match our tone, cite current policy docs, and run cheaply at scale, which combination of prompting, RAG, and fine-tuning would you use and why?

References

• Hugging Face — PEFT documentation
• OpenAI — When to fine-tune (guidance)