NGM-RAG: Neural Graph Matching based Retrieval-Augmented Generation

Traditional Retrieval-Augmented Generation (RAG) systems often struggle with complex, multi-hop questions because they rely primarily on text-based retrieval, which fails to capture the rich, relational connections between entities. The authors address this by proposing NGM-RAG, a framework that explicitly incorporates graph structures into the retrieval process to improve the accuracy of context selection for LLMs.

NGM-RAG introduces a unified process consisting of three key stages: graph construction, graph matching, and answer generation.

• Graph Construction: The system uses an LLM to extract entities and relationships from the corpus to build a knowledge graph.
• Graph Matching: This is the core innovation. Instead of relying solely on text similarity, the system employs a hybrid matching strategy. It combines traditional methods (Levenshtein distance for names and BM25 for text descriptions) with a Neural Graph Matching approach. This neural component uses Graph Neural Networks (GNNs)—specifically LightGCN for homogeneous graphs and GINE for heterogeneous graphs—to compute matching scores based on graph topology.
• Adaptive Weighting: The framework uses an adaptive weighting strategy to balance these different matching signals, ensuring the most relevant nodes are selected for the final generation step.

The authors evaluated NGM-RAG on multi-hop question answering (HotpotQA, MultiHop-RAG) and long-context summarization (UltraDomain). The results demonstrate that NGM-RAG consistently outperforms traditional NaiveRAG and state-of-the-art graph-enhanced baselines like GraphRAG and LightRAG. The model shows robust performance across various backbone LLMs, including Llama-3.1-8B and DeepSeek-R1, suggesting that the neural graph matching approach effectively bridges the gap between structured relational data and LLM reasoning capabilities.