I saw another repo use weights from llama3 to construct embedding model. I think the use case will be use this small model to search and use bigger model to re-rank later. So the question will be how this approach compared to BM25