E$^2$GraphRAG: Advancing the Pareto Frontier in Efficiency and Effectiveness for Graph-based RAG

ICLR 2026 Conference Submission17284 Authors

Keywords: Efficient RAG, Graph-based RAG

TL;DR: We propose E^2GraphRAG, a streamlined graph-based RAG framework that advances the Pareto frontier with respect to efficiency and effectiveness.

Abstract: Graph-based RAG methods like GraphRAG demonstrate strong global understanding of the knowledge base by constructing hierarchical entity graphs, but often suffer from inefficiency and rigid, manually defined query modes, limiting practical use. To address these limitations, we present E$^2$GraphRAG, a streamlined graph-based RAG framework that advances the Pareto frontier of Efficiency and Effectiveness. In the indexing stage, E$^2$GraphRAG utilizes large language models to generate a summary tree, and NLP tools to construct an entity graph from document chunks, with bidirectional indexes linking entities and chunks for efficient lookup. In the retrieval stage, the graph structure filters related entities, while the bidirectional indexes map these entities to their corresponding chunks, supporting an adaptive mechanism that dynamically switches between local and global modes. Experiments show that E$^2$GraphRAG achieves up to $10\times$ faster indexing than GraphRAG while maintaining comparable QA performance, advancing the Pareto frontier with respect to effectiveness and efficiency. Our code is available at https://anonymous.4open.science/r/E-2GraphRAG-8897.

Primary Area: generative models

Submission Number: 17284