If the self evaluation makes it better, then why not do the self evaluation as part of the normal RAG workflow?