Prompting, RAG or Fine-tuning

You have three main ways to make a model work better for your needs. Better prompts, retrieval, and fine tuning. Beginners often jump to the hardest one first. In this lesson, you will learn a simple order of attack, that saves time, and money.

• Three ways to customise
• Start with the cheapest
• Move up only when needed

The ladder of customisation

Think of a ladder. The bottom step is prompting. It is instant, free, and surprisingly powerful. If the model lacks knowledge, climb to retrieval, and give it your documents. If it still does not behave, in the way you need, in its style, its format, or its skill, then consider fine tuning. Each step costs more time, and money, so only climb when the step below is not enough.

1. 1. Promptinginstant and free: try this first
2. 2. Retrieval (RAG)give it your documents and facts
3. 3. Fine-tuningteach a new style or skill
4. 4. Train your ownalmost never needed

What each method is good at

Each method fixes a different problem. Prompting fixes unclear instructions. Retrieval fixes missing, or changing knowledge, like policies and product details. Fine tuning fixes behaviour, a consistent tone, a strict format, or a specialised skill. A common mistake is fine tuning, to add facts. Facts change, and are better retrieved. Behaviour is stable, and is better learned.

Real examples

A few examples. A shop wants answers about its return policy. That is retrieval. A law firm wants every summary, in one exact house style. That is fine tuning, perhaps after strong prompting. A developer wants a model to follow a rare internal format, reliably. Fine tuning again. And a student wants help with homework. That is just good prompting, and nothing else.

Summary