Disco-RAG Framework Enhances AI Knowledge Synthesis with Discourse-Aware Retrieval

A new AI framework called Disco-RAG, detailed in arXiv preprint 2601.04377v5, addresses limitations in Retrieval-Augmented Generation (RAG) by incorporating discourse structures to improve knowledge synthesis. Traditional RAG methods often handle retrieved passages in a flat, unstructured manner, which hinders the capture of structural cues and the integration of evidence across multiple documents. Disco-RAG overcomes this by explicitly injecting discourse signals into the generation process. It constructs intra-chunk discourse trees to model local hierarchies and builds inter-chunk rhetorical graphs to represent cross-passage coherence. These structures are combined into a planning blueprint that conditions the generation output. Experiments on benchmarks for question answering and long-document summarization demonstrate the framework's effectiveness, showing improved performance in knowledge-intensive tasks. The approach aims to enhance large language models (LLMs) by enabling better synthesis of dispersed information from various documents. This development is part of ongoing research in AI and machine learning, focusing on advanced retrieval and generation techniques. The preprint was announced as a replacement cross on arXiv, indicating an update to prior versions.