Fine-tuning a small model for Nepali customer support — when is it worth it vs just RAG?

Spent 3 weeks evaluating fine-tuning a 7B model on 4000 support tickets vs a simpler RAG over the same data.

Result: RAG won on accuracy. Fine-tuned model was faster but kept making up product names that didn’t exist.

Fine-tuning is worth it for STYLE (tone, format consistency). RAG is worth it for FACTS. Stop conflating the two.