Is privateGPT what you're looking for? It stores PDFs in a vector DB and answers questions based on that content alone.
Edit: Everything is local. This is not training, it takes a standard LLM and gives it extra context based on what text the tool finds in the vector DB related to the topic.