This is an interesting feature. But you don't need to use just OpenAI's embeddings. You can generate your own embeddings with open source SOTA transformer models which would probably work just the same. You could generate a couple hundred thousand embeddings with a rented A100 for less than 2 dollars. And the point of converting text or other objects (like images) into embeddings is to compare a large number of documents to a source document very fast. It's more useful to put the embeddings in something like Redis. This pgvector data type would be good for an offline backup of vectors.