Retrieval-Augmented Generation (RAG) enhances the answer quality of language models by incorporating external knowledge. However, most RAG systems do not clearly explain how evidence is retrieved, and can be slow with large-scale databases. This paper proposes a novel graph-based RAG method that selects landmark nodes using K-means clustering and constructs a sparse graph for efficient and interpretable retrieval. Text passages are embedded using Sentence-BERT, and the centroid of each cluster serves as a landmark. During inference, queries are mapped to the nearest landmark, and relevant evidence is identified by traversing the landmark graph linked to stored embeddings. The proposed approach enables users to understand how evidence is retrieved within RAG, and provides richer input information for the prompt, which in turn improves the answer quality of generative AI models.