ChatWeaver: Interactive Knowledge Graph for Efficient Information Retrieval and Visualization Empowered by LLM

The verbose responses generated by Large Language Models during multi-turn interactions substantially impair reading efficiency, while their limited short-term memory mechanisms hinder users' ability to perform depth-breadth exploratory analysis on historical dialogues. To address these dual challenges, our study proposes a collaborative framework that integrates explicit knowledge bases with implicit model knowledge through an interactive knowledge graph, achieving three innovations: (1) A dual-source knowledge repository (integrating documents and conversation histories) enables dynamic editing, employing incremental label propagation algorithms to extract community semantic summaries and construct hierarchical indices, while supporting both depth-first (DFS) and breadth-first (BFS) exploration pathways; (2) A retrieval logic visualization system explicitly maps key entity-relationship attributes into interactive graph representations, enhancing dynamic parsing of complex relationships and interpretability of search results; (3) A user-driven knowledge maintenance interface permits real-time node correction, relationship restructuring, and semantic label expansion, establishing human-AI collaborative knowledge evolution. By organically integrating structured knowledge editing with dynamic and intuitive exploration, our solution demonstrates a significant improvement in the accuracy, efficiency, and user engagement of knowledge retrieval, outperforming traditional methods in various benchmark tests.