[rolisz]

Up until a month ago, the OpenAI embeddings where very poor. But they recently released a new model which is much better then they're previous one.

Now it depends un specific usecase (domain, language, length of texts)