Fine-tuning

The process of continuing to train a base model with examples specific to your domain so it can learn a particular style, format, or vocabulary. It requires a curated dataset (typically 500–5,000 examples), training infrastructure, and technical expertise. The resulting model is “yours,” but it still inherits the capabilities and biases of the base model.

When should I use Fine-tuning at a company?

When neither prompting nor RAG achieves the required consistency in style or format; when you need highly specific technical vocabulary; when latency or cost per inference matters more than flexibility.

When is it NOT a good idea Fine-tuning?

As a first option. Always start with prompting and scale to RAG; fine-tune only when the first two aren’t enough. The typical investment is 10–30x higher than an equivalent RAG.