Ah, but this is nice project. I'll start hacking once it's easier to fine-tune it with own documents for specific questions. What plaques me, though, is how you prevent the model from answering questions it was not trained for?