RAFT:  A new way to teach LLMs to be better at RAG

I understand this approach, just add some useful information of filenames (just filename) in train data. But when i got some new file in my rag database, this approach highly can not work well. I think this approach can't solve both limits of fine-tune and rag, just temporarily reinforced model generation.  Why not fine-tune embedding or rerank model?