We just shipped a healthcare document assistant that uses both RAG and fine-tuning, and the biggest lesson was this: they solve completely different problems. RAG keeps your system truthful (retrieves fresh docs at query time). Fine-tuning keeps it consistent (encodes behavior into weights). Trying to force one to do both jobs is where most teams waste months. A few things that surprised us:

• For knowledge bases under 200K tokens, you can skip RAG entirely and just load everything into the context window with prompt caching. Way simpler.
• Fine-tuning's real cost isn't GPU time, it's data curation. That part took 3x longer than we estimated.
• The 2025 LaRA benchmark (ICML) confirmed there's no universal winner. Hybrid is the production default now.

Wrote up the full decision framework with a comparison table and practical examples here: adamosoftware.hashnode.dev/rag-vs-fine-tuning-in-…

For anyone building AI features right now: are you going RAG-first, fine-tuning-first, or hybrid from day one?