A Graph Retrieval-Augmented Generation Framework for AI-Powered Supplier Discovery

Abstract: As modern supply chains demand higher resilience, agility, and visibility, the task of supplier discovery becomes increasingly critical. We present a novel AI-powered methodology that combines Graph Neural Networks (GNNs), Retrieval-Augmented Generation (RAG), and Large Language Models (LLMs) to enhance supplier search and reasoning. A structured Supplier Capability Knowledge Graph (SCKG) is built by extracting domain-specific triplets from unstructured manufacturing data using fine-tuned LLMs and is enriched through semantic normalization via ontology and manufacturing thesaurus. A GNN-based retrieval system identifies relevant subgraphs by performing dense reasoning over the SCKG. These subgraphs are verbalized into natural language using shortest-path reasoning chains and fed into an LLM for generative explanation. To improve retrieval precision, a hybrid entity normalization technique leveraging Jaccard similarity and vector-based retrieval is applied. This integrated GNN-RAG system significantly outperforms traditional and zero-shot LLM-based supplier search approaches in both precision and recall on real-world datasets. Our results demonstrate the system's ability to perform robust, real-time supplier discovery while enabling explainable and accurate responses.

Keywords: Supplier Discovery, Knowledge Graphs, GNN-RAG, Large Language Models, Semantic NormalizationI