CAGE: Coherence-Aware Graph Encoding for Retrieval-Augmented Generation

The researchers proposed the CAGE (Coherence-Aware Graph Encoding) framework, aimed at addressing the issue of lack of overall consistency in the context sets of traditional RAG (Retrieval-Augmented Generation) systems. This framework models “block consistency” in four dimensions: relevance, noise resistance, information connectivity, and factual consistency, transforming retrieval paragraphs into directed graphs. By weighting fact anchors based on minimum outdegrees and using graph convolutional networks to encode structural patterns, CAGE achieves a fusion of block consistency with query relevance in sorting. In four multi-hop benchmark tests, CAGE achieved comparable or better recall@5 performance on bridge-dominated datasets compared to strong baselines like monoT5, while continuously improving the downstream precise matching rate. This demonstrates that context with structural consistency can provide more accurate answers even when the recall rate is low.