From Local to Global: A Graph RAG Approach to Query-Focused Summarization
Topics: AI (Deep Learning), Graph RAG, Knowledge Graph, Microsoft, Retrieval Augmented Generation (RAG)
The paper “From Local to Global: A Graph RAG Approach to Query-Focused Summarization” presents a new method that enhances the capabilities of large language models (LLMs) for query-focused summarization (QFS) tasks over large text corpora. By creating an entity knowledge graph and using community detection algorithms, the approach partitions the graph into communities, enabling the generation of comprehensive and diverse answers to global sensemaking questions. Evaluations show that this Graph RAG method significantly outperforms traditional RAG approaches in terms of comprehensiveness and diversity of answers.