How can fine-tuning embedding models improve the retriever's performance in RAG?

How can fine-tuning embedding models improve the retriever's performance in RAG?