Retrieval-augmented generation (RAG) systems have demonstrated potential in enhancing large language models with external knowledge sources, yet they face significant limitations when handling multiple questions requiring complex reasoning across multiple documents. This paper introduces Multi-Hop Graph Attention RAG (MHGA-RAG), a novel framework that synergistically integrates knowledge graph-based retrieval with attention-enhanced mechanisms to address these challenges. Our approach uniquely combines structured knowledge representation with semantic context modeling, creating a comprehensive solution for multi-hop reasoning tasks. MHGA-RAG incorporates three key innovations: (1) a knowledge graph construction methodology that captures entity relationships and reasoning paths, (2) a graph-augmented retrieval component that leverages co-occurrence patterns to expand queries contextually, and (3) an attention-enhanced mechanism that dynamically balances semantic, relational, and structural features through multi-head attention. Extensive experiments on both domain-specific (Crinoid Music Robot) and general knowledge (HotpotQA) datasets demonstrate that MHGA-RAG significantly outperforms traditional RAG methods in response quality, information richness, and accuracy, while maintaining strong cross-domain generalization capabilities. Ablation studies confirm the complementary value of integrating knowledge graph structures with attention mechanisms, particularly for information-dense scenarios requiring multi-hop reasoning.