Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by grounding them in external knowledge, yet most methods retrieve isolated text chunks, overlooking critical interrelationships. Recent graph-based RAG approaches attempt to model these connections but often produce simple factual chains, missing the dense thematic clusters vital for complex reasoning. In this paper, we introduce KGCRAG, a novel framework that employs community detection to address this limitation, extracting cohesive clusters from the knowledge graph to form a focused, relevant context for the LLM. Featuring a timeout-driven fallback mechanism, our approach ensures efficiency and reliability. Extensive experiments on standard benchmarks (HotpotQA, MuSiQue) and our newly constructed academic QA dataset (ScholatQA) demonstrate that KGCRAG outperforms existing RAG methods in response accuracy and efficiency.