Interesting. What is your LLM prompt to achieve that?

Have you tried using text transformer embedding model and use vector distance to measure similarity for spam rate limit?