it really can't be that difficult to have a llm reference a known factual reference before giving an answer. Its really good at figuring out what you want and what to say, not far off to check references
Think for a minute about how undefined and vague what you are asking is. Unless the specific claim you have made is close to verbatim in some reference (in which case you're doing search & retrieval), it's incredibly hard to "automate" the interpretation of meaning from the source doc, your query, and the token being generated by the LLM -- such that you can give a sensible citation. There's a long tail of cases where this will break, and the moment you expose this functionality to users they will have the easy cases for breakfast, quickly start exploring the long tail and complain about failures.
