The tooling available to train a model is easy. I started that day having no hands-on exposure. I ended the day with a minimal RAG that could provide answers related to the questions I was asking. the answers were NEAR the question, but not directly answering it. The tooling for a self-hoseted model is not difficult to use and is getting better. The learning curve to me seems to be in how to best insert the data into the vector DB to get more accurate answers. How it's sliced matters and I'm learning about that right now with an internal POC.