← All lab updates
Prompt, retrieval or fine-tune: which one fixes which problem
Most teams reach for a longer prompt whatever the problem is. A prompt changes what you ask, retrieval changes what the model knows at answer time, fine-tuning changes how it behaves. A short guide to telling the three apart, with a table you can use today.
## Three tools, three different jobs
- **The prompt** changes what you ask. It is cheap and instant, and it is also the first thing to wear off in a long conversation.
- **Retrieval** (often called RAG, retrieval-augmented generation) hands the model the right documents at the moment it answers. It fixes missing or stale knowledge. It does not change habits.
- **Fine-tuning** changes the model's weights. It fixes behaviour that keeps coming back: a tone, a format, a way of giving in.
## A quick table
| The problem | Try first |
|---|---|
| The model does not know your product or your policies | Retrieval |
| The facts change every week | Retrieval |
| Wrong format or tone, every time | Prompt, then fine-tune if it does not hold |
| It gives in when the user pushes back | Fine-tune |
| It obeys instructions hidden in a document or email it reads | Fine-tune, and read the model on that boundary |
| It invents sources | Retrieval for the facts, fine-tune for the habit of inventing |
## Why behaviour resists prompts
A system prompt is one message at the top of the conversation. Every turn after it adds pressure from the other side: the user insists, flatters, sets a deadline. Over ten or twenty turns the instruction loses weight. A habit that lives in the weights does not have to compete for attention in the same way, which is why we correct drift in the model rather than by writing longer prompts.
## A good order of work
1. Write the rule down clearly in the prompt, and measure.
2. If the problem is knowledge, add retrieval, and measure again.
3. If the problem is behaviour and it survives both, fine-tune on examples of the right behaviour, and check that nothing else got worse.
Measure each step on the conversations your users really have, not on a demo.
## Go further
- [RAG and retrieval](https://xbe.at/llms-and-genai/topic/rag-and-retrieval/) and [Fine-tuning and alignment](https://xbe.at/llms-and-genai/topic/fine-tuning-and-alignment/), practical articles on xbe.at
- [How we fine-tune a model](/fine-tuning) and [The layers we put on a model](/model-layers)
This update is not available.