I like this approach, just a bit too expensive for large datasets and local LLMs not that great and harder to scale, this needs to mature a bit more but gets my thumbs up.