Oh yeah, that looks super similar. I remember the similarity score being tricky to get useful signal out of, for the underlying model I had used back then. Similar and dissimilar issues all hovered around the 0.80 mark. But surely not hard to improve on, with larger models and possibly higher-dimension vectors.