The precision is around 60% - 70%. It could be increased with LLMs+RAG and I've tried it already, but that would cost some $$ on this amount of comments.