TGS-RAG: Bidirectional Text-Graph Framework for LLM Retrieval

A novel framework named TGS-RAG (Text-Graph Synergistic enhancement) has been introduced to tackle the 'Information Island' challenge in Retrieval-Augmented Generation (RAG) for Large Language Models (LLMs). Conventional text-based RAG often retrieves irrelevant pseudo-evidence, while graph-based RAG is hindered by search-time pruning that eliminates valid reasoning paths. Current hybrid methods rely on basic evidence concatenation or one-way enhancement, which do not adequately address the asymmetric reasoning dynamics between unstructured text and structured graphs. TGS-RAG features a bidirectional approach: a Graph-to-Text channel that utilizes a Global Voting mechanism from engaged graph nodes to re-rank and enhance textual evidence, effectively reducing semantic noise. This framework seeks to improve factual grounding and multi-hop reasoning in LLMs. The research can be found on arXiv under ID 2605.05643.