RAG vs. GraphRAG: A Systematic Evaluation and Key Insights

This paper provides a systematic comparison between traditional RAG and GraphRAG. GraphRAG achieves 86.31% accuracy on RobustQA benchmarks, compared to approximately 28% for baseline RAG. GraphRAG excels on multi-hop reasoning, thematic queries, and relationship understanding. Traditional RAG still performs well on direct fact-lookup queries and latency-sensitive applications.