What you're describing is just called "question answering" in NLP afaik. A language model will take in a source document and a question and spit out either a generated answer to the question or a section of the source text containing the answer.

Check some of the QA models on huggingface to get an idea if you're not already familiar