Curious how semantic cache works,are cache hit based on embeddings, exact prompt normalization, how do you avoid serving stale incorrect cache..