Question about returning document references to build answer.

[digredeth]

I'm building a simple app that scans a couple of documents and then gets vectorized. From the UI select which document want to ask questions and then I send it. The thing is that documents are large (between 70 and 200 pages) so I'm using loadQAMapReduceChain to get context from the whole document avoiding language model token restrictions.

The question is, what would be a good approach to also be able to return the pages of the original document used to build the question answer? Pages numbers are not available in the document text. Should I attach it to intermediate chunks? should I add it as chunk metadata?