They're not using vector embeddings for determining similarity - they use finetuned NLI models that take the context into account to determine semantic equivale

They're not using vector embeddings for determining similarity - they use finetuned NLI models that take the context into account to determine semantic equivalence. So it doesn't depend on knowing properties of the output distribution up front at all. All you need to be able to do is draw a representative sample (up to your preferred error bounds).