Fine-tuning
Fine-tuning further trains a base model on labelled examples so behaviour fits a narrower domain or format.
How it works
Collect high-quality examples, train, then evaluate on held-out cases. Prefer prompts and RAG first; fine-tune when eval proves a persistent gap that retrieval cannot fix.
Example
A team fine-tunes for a strict JSON report format after prompt-only attempts fail the schema eval set.