Innovative Framework for Enhanced Textual Graph Understanding

The integration of textual graph-based retrieval-augmented generation (GraphRAG) has proven to be an effective strategy for refining large language models (LLMs) in specialized question answering tasks. Traditional methods have largely concentrated on zero-shot GraphRAG, but the selection of high-quality demonstrations is essential for enhancing reasoning capabilities and accuracy in responses. Recent findings indicate that the subgraphs retrieved can often include extraneous information, negatively impacting reasoning efficiency. In response to these challenges, we present MixDemo, an innovative GraphRAG framework that employs a Mixture-of-Experts (MoE) system to identify the most pertinent demonstrations tailored to varying question contexts. To minimize irrelevant data within the retrieved subgraphs, we also introduce a query-specific graph encoder designed to focus on the most relevant information for each query. Comprehensive testing across various textual graph benchmarks demonstrates that MixDemo significantly surpasses current methodologies.