What is fine-tuning?

A large language model arrives trained on a vast amount of general text. Fine-tuning takes that model and trains it a little more on examples you choose: hundreds or thousands of inputs, each paired with the output you want. The model adjusts, and from then on it leans towards that style, format or judgement without being told each time.

An example. A distributor receives thousands of supplier emails a week and wants each one sorted into the right queue: price change, delivery delay, product withdrawn, query. Fine-tuning a model on past emails that the team has already sorted teaches it the distributor's own categories and the signs of each, so it sorts new emails the same way, quickly and consistently.

Fine-tuning and RAG

The two are often confused, and they do different jobs. Fine-tuning changes how a model writes and behaves. RAG changes what it knows at the moment it answers, by looking things up in your documents and data first.

• Use RAG when the answer depends on information that changes: products, prices, stock, policies, a customer's account. Update the source and the next answer uses it.
• Use fine-tuning when you need a consistent style, format or decision that's hard to describe in instructions: your tone of voice, a fixed output structure, your own categories.
• Use both when you need both: a model fine-tuned to write in your format, answering from your current information through RAG.

For answering from business information that changes, RAG is usually the right starting point. Fine-tuning can also let a smaller, faster model do a narrow job well, which can bring down the cost of running it.

What it needs to do well

Good fine-tuning starts with good examples: enough of them, correct, consistent and representative of the real work. It also needs a set of test cases kept back, so you can measure whether the tuned model is better than the original, and a plan to tune again when the base model or your needs change.

How we approach it

We start from the job, not the technique. Whether it calls for RAG, fine-tuning, a specialised model or plain code, the work begins with getting your data into the shape the AI can use.