Systematic Knowledge Injection into Large Language Models via Diverse Augmentation for Domain-Specific RAG