A review on Graph RAG methods

Large Language Models (LLMs), built upon the transformer architecture and capable of few-shot learning, are nevertheless constrained by their static, pre-trained knowledge and a propensity for hallucination. Retrieval-Augmented Generation (RAG) addresses these issues by grounding generation in external data. However, standard ``Naive RAG'' frameworks, often relying on dense passage retrieval, sequence-to-sequence denoising, and vector stores like Chroma, fail when handling complex, relational data, as they treat knowledge as an unstructured ``bag of chunks.'' To resolve this, a new paradigm reviewed by, Graph RAG, has emerged, transforming source corpora into structured knowledge graphs. While advanced frameworks like ``From Local to Global'' (GraphRAG), StructRAG, and LightRAG claim superiority over retrieval-less approaches like HOLMES or graph optimization methods like REANO, they have been developed in isolation. This creates a critical research gap, leaving no clear, comparative data for practitioners. This thesis bridges this gap by presenting a rigorous benchmark analysis of these leading Graph RAG frameworks using the Gemini 2.0 and 2.5 flash-lite models as the underlying reasoning engines. They are evaluated against Naive RAG and a Simple LLM baseline across two distinct datasets: UltraDomainCS and NarrativeQA. Performance is measured using a broad range of metrics, including quality via LLM-as-a-Judge and BERTScore, cost, latency, and robustness. Our findings quantify the limitations of unstructured retrieval and the specific trade-offs of graph-based approaches. On technical data, Naive RAG proved detrimental (43.82\%), inferior to the Simple LLM baseline (49.20\%). In contrast, LightRAG emerged as the optimal solution (92.75\%) while consuming approximately 8x fewer tokens than frameworks like StructRAG. Conversely, on narrative data, StructRAG was the only robust performer, uniquely capable of bridging long-distance relationships, though at high cost. Finally, while GraphRAG provided high-quality answers, its latency rendered it non-viable for real-time applications. This analysis establishes the first data-driven trade-off for Graph RAG, demonstrating that the choice of framework must be optimized for the specific density and cost constraints of the domain.