[quie]

Why not implement check_answer by feeding player answer back into LLM and ask (and even explain) whether it is correct (or not)? This way there is no need to have any explicitly coded similarity logic.