I think the issue is LLMs don't scale well to the level needed for search. They also have obvious latency. So generative usage seems to be the best fit at the moment.
But you could feed the text through the LLM, generate a summary and a list of keywords, characters and locations and search through that the regular way. And as far as I can tell, that doesn't really exist yet.
IMDb has a keyword feature[1], but that's still completely human-curated and quite useless due to that, as it just doesn't have the necessary depths and completeness to discover interesting new content.
IMDb has a keyword feature[1], but that's still completely human-curated and quite useless due to that, as it just doesn't have the necessary depths and completeness to discover interesting new content.
[1] https://www.imdb.com/search/keyword/