I think a more robust approach would be to restrict the generative AI to generating summaries of book texts. First summarize every book (this only has to be don

I think a more robust approach would be to restrict the generative AI to generating summaries of book texts. First summarize every book (this only has to be done once), and then use vector search to find the most similar summaries to the provided summary. Small mistakes will make little difference, e.g. "goblin" will have a similar embedding to "demon", and even entirely wrong information will only increase the number of books that have to be manually checked. Or better yet, develop an embedding model that can handle whole books at once and compare the vectors directly. Perhaps somebody with more compute than they know what to do with could try this with Project Gutenberg as a proof of concept.