How do the categories work? I'd download the video transcripts and embed the text (maybe summarized first). This way you could support custom categories via embedding similarity.