Coordinated Semantic Alignment and Evidence Constraints for Retrieval-Augmented Generation with Large Language Models — arXiv2