It could still remember where each spell is mentioned. I think the only way to properly test this would be to run it against an unpublished manuscript.
If you ask a model to discuss an obscure work it'll have no clue what it's about.
This is very different than asking about Harry Potter.