I think he might be saying, have metadata in your vector retrieval that describe the domain of the retrieved chunk and use that as a decision on which model to use downstream. Sounds like very interesting improvement of RAG