Module 10 · Lesson 1 of 2
Prompt, retrieve or fine-tune?
Three ways to improve a model
A better prompt changes what you ask, retrieval changes what the model reads, and fine-tuning changes how it behaves.
Scene 1 of 5 · Ask better first
Clearer instructions and two or three worked examples often fix a weak answer for almost nothing.
- Your task — healthy
- Better prompt — healthy — what you ask
- Retrieval — healthy
- Fine-tuning — healthy
| Option | Changes | Cost |
|---|---|---|
| Better prompt (highlighted) | What you ask | Lowest |
| Retrieval | What it can read | Medium |
| Fine-tuning | How it behaves | Highest |
- Ask better first. Clearer instructions and two or three worked examples often fix a weak answer for almost nothing.
- Let it read. When answers need private or changing facts, retrieval puts the current text in front of the model.
- Train its habits. Fine-tuning reshapes behaviour, such as one narrow task done cheaply by a smaller model.
- Facts go stale. Facts trained into a model stay frozen, so a price change would mean training all over again.
- Climb only when needed. Each option costs more than the last, so a test set decides when to climb.
- Now you try. Pick a need, and see which option to reach for first. For example: Start with retrieval. The facts change, so the model must read the current text at question time.
Scene 1 of 5: Ask better first. Clearer instructions and two or three worked examples often fix a weak answer for almost nothing.
Key points
- Try a better prompt first: it costs least to change and test.
- Facts that change or are private call for retrieval.
- A narrow task at high volume can justify fine-tuning a smaller model.
- A test set decides whether each step paid off.
There are three main ways to make a model better at your task. Change the (Everything sent to a language model in one request: instructions, context and the question.), with clearer instructions and a few worked examples. Retrieve your documents into the prompt at question time. Or (Training an already trained model a little more on your own examples, to change how it behaves on your task.): train the model further on your own examples.
Start with the prompt, since it is the cheapest to change and test. Add retrieval when answers need facts the model never saw, or facts that change, such as prices or policies.
Fine-tuning changes how a model behaves: a fixed format, a tone, or one narrow task done well, often by a smaller and cheaper model. It is a poor way to add facts that change, since every change needs another round of (The long, costly process in which a model learns from data. It happens before the model is used, not during each request.).
Each option costs more to build and run than the one before. Measure each against your test set, and move on only when the cheaper option falls short.
In system design: Prompt first, retrieval for changing facts, fine-tuning for fixed behaviour at scale. Each step needs a test set to show it helped.