GraphRAG: Reasoning on Graphs with Retrieval-Augmented LLMs

Recently, large language models (LLMs) have made significant advancements towards various tasks in natural language processing. However, LLMs often lack knowledge about data that is internal to enterprises, some of which is structured as graphs, and it remains unclear how to enable LLMs to reason on external graph-structured data. To address this, we propose a retrieval-augmented generative model, \textit{GraphRAG}, that employs a structure-aware retriever to leverage information from graph-structured data and leverages an LLM to reason on the retrieved structural data. We evaluate the model on the task of predicting product categories of an e-commerce portal that are relevant to a webpage, when the categories are organized as a collection of disjoint trees. Extensive experiments show that GraphRAG improves 10.22% in precision@3 over baseline retrieval-augmented generation models.