The increasing deployment of Retrieval-Augmented Generation (RAG) systems in critical domains necessitates robust mechanisms for tracing information sources in AI-generated content. This article presents a novel approach to fine-grained source attribution in RAG-powered Artificial General Intelligence (AGI) responses through a multi-stage architecture combining Source-Preserving Embedding (SPE) and Source-Aware Attention (SAA) mechanisms. Our system employs a modified T5 architecture with 3.2 billion parameters and a graph-based SourceRank algorithm for post-generation attribution analysis. Evaluated across 10,000 queries in five domains (medicine, law, finance, technology, and general knowledge), the system achieved 87.3% attribution accuracy, significantly improving baseline methods. The article demonstrated particular strength in handling domain-specific content, maintaining high precision-recall balance, and managing complex multi-source attributions. User studies with 50 domain experts validated the system's effectiveness, with 92% expert agreement on attributions and 88% rating the attribution information as highly helpful. While computational overhead and multi-hop reasoning scenarios present ongoing challenges, our approach significantly advances the transparency and trustworthiness of RAG-powered AGI systems.