Generally if you are looking to provide additional knowledge to GPT the best way is through embeddings, not fine-tuning (which is better at classification/structure changes). LangChain project is probably the easiest way to get a proof-of-concept running so you can evaluate the results.
The fine tuning guide I linked has all the information on that format.
Again, the consensus is not to use fine tuning for knowledgebase Q&A chatbot style applications. Here’s a relevant LangChain tutorial. You can search these forums for fine-tuning vs. embeddings for more discussion.