Fine-tuning
Fine-tuning means taking an already trained language model (a general-purpose LLM) and partially retraining it on a dataset specific to a domain, tone, or task. The model keeps its general abilities, but aligns better with the company's vocabulary, formats, and use cases.
It becomes relevant when prompt engineering and RAG are no longer enough: a tightly constrained writing style, complex business classification, or an output format that must stay stable at high volume. On a product-sheet generation project, for example, fine-tuning on a few thousand validated examples can yield a more consistent tone and structure than a long prompt. It's a heavier investment than RAG, reserved for cases where the quality or cost gain justifies it.
The classic trap is launching fine-tuning too early, when a solid prompt and a well-indexed document base would have been enough. Fine-tuning costs time (data prep, training, evaluation) and freezes the model: every change in business rules may require retraining. Start with prompting and RAG, and fine-tune only once limits are measured, not imagined.
Follow Fragments Studio on Google
Add us to your preferred sources and our articles get surfaced first in Top Stories, AI Overviews and AI Mode.
