Usually fit those chatbots you take a trained model and use RAG, essentially turning the question into a traditional search and asking the LLM to summarize the contents from the result. So it’s frequently a convenient front end to a search engine, which is how it avoid s having to train to produce relevant responses. Is generally just prohibitively difficult in various ways to fine tune LLM through training and manage to get the desired behavior. So it can act like it “knows” about the stuff you do despite zero training if other methods are stuffing the prompts with the right answers.