I'm really curious to see where that investigation leads - have you done any comparisons between Llama 2 and the embedding focused models? I wonder if it'll be better able to provide more 'intuitively correct' similarities?