Could you try using a built-in model like s-bert, instead of OpenAI, to see if that also has the same issue?