Comment by mdagostino

I wouldn’t say most—maybe a factor of 2. Getting the embedding is still an API call to an LLM.