This uses the OpenAI API currently. I tried Gemini for embeddings for a little while but didn't seem materially better.