Reproduce “FRAG: A Flexible Modular Framework for Retrieval-Augmented Generation based on Knowledge Graphs” on OpenResearch

About this paper

Zengyi Gao, Yukun Cao, Hairu Wang, Ao Ke, Yuan Feng, Xike Xie, S Kevin Zhou

The FRAG framework introduces a flexible and modular approach to Retrieval-Augmented Generation (RAG) using Knowledge Graphs (KGs) to enhance Large Language Model (LLM) accuracy. It achieves state-of-the-art performance among modular KG-RAG methods by adaptively tailoring knowledge retrieval based on query complexity, significantly outperforming prior methods like ToG and achieving competitive results with coupled approaches at a fraction of the computational cost.

Key insights

• Query complexity, discernible from the query text itself, can be effectively leveraged to guide and adapt knowledge graph retrieval strategies.
• It is possible to achieve high-quality, adaptive knowledge graph retrieval within a modular framework without requiring extensive LLM fine-tuning or iterative LLM calls during the retrieval phase.
• Tailoring subgraph extraction and path-finding algorithms (e.g., BFS for simple, Dijkstra for complex) based on inferred query characteristics significantly improves retrieval efficiency and relevance.

Summary by alphaXiv.