Traditional RAG has a semantic segmentation problem where mechanically splitting content into fixed-length chunks often cuts off complete logical units, resulting in incomplete context in retrieved fragments.