Abstract:Large Language Models~(LLMs) have demonstrated capabilities across various applications but face challenges such as hallucination, limited reasoning abilities, and factual inconsistencies, especially when tackling complex, domain-specific tasks like question answering~(QA). While Knowledge Graphs~(KGs) have been shown to help mitigate these issues, research on the integration of LLMs with background KGs remains limited. In particular, user accessibility and the flexibility of the underlying KG have not been thoroughly explored. We introduce AGENTiGraph (Adaptive Generative ENgine for Task-based Interaction and Graphical Representation), a platform for knowledge management through natural language interaction. It integrates knowledge extraction, integration, and real-time visualization. AGENTiGraph employs a multi-agent architecture to dynamically interpret user intents, manage tasks, and integrate new knowledge, ensuring adaptability to evolving user requirements and data contexts. Our approach demonstrates superior performance in knowledge graph interactions, particularly for complex domain-specific tasks. Experimental results on a dataset of 3,500 test cases show AGENTiGraph significantly outperforms state-of-the-art zero-shot baselines, achieving 95.12\% accuracy in task classification and 90.45\% success rate in task execution. User studies corroborate its effectiveness in real-world scenarios. To showcase versatility, we extended AGENTiGraph to legislation and healthcare domains, constructing specialized KGs capable of answering complex queries in legal and medical contexts.

What problem does this paper attempt to address?

The problems that this paper attempts to solve mainly focus on the following aspects: 1. **Limitations of large - language models (LLMs) in question - answering tasks**: - **Hallucination problem**: LLMs sometimes generate information that does not conform to the facts. - **Limited reasoning ability**: LLMs have insufficient reasoning ability when dealing with complex, domain - specific tasks. - **Factual inconsistency**: When providing answers, LLMs may have factual inconsistencies. 2. **Integration problems between knowledge graphs (KGs) and LLMs**: - **User accessibility**: Although existing KG tools and query languages (such as SPARQL and Cypher) are powerful, they lack user - friendly interfaces and require professional technical knowledge, which limits their wide use. - **Lack of flexibility**: Existing systems often lack context understanding and flexibility when dealing with complex or subtle queries. - **Seamless integration**: Current tools have difficulties in seamlessly integrating KGs with natural - language interfaces. 3. **Processing of multi - modal inputs**: - Existing systems lack effective mechanisms to create coherent action plans when dealing with multi - modal inputs (such as user intentions, query histories, and graph structures). To address these challenges, the paper proposes the **AGENTiGraph** platform, which solves the above problems through the following innovations: - **Semantic parsing**: Optimize user interactions, convert natural - language queries into structured graph operations, and improve the accuracy and speed of request processing. - **Adaptive multi - agent system**: Integrate multi - modal inputs, dynamically interpret user intentions, manage tasks, and integrate new knowledge to ensure adaptation to changing user needs and data environments. - **Dynamic knowledge integration**: Support continuous knowledge extraction and integration, ensure that the knowledge graph is up - to - date, and provide dynamic visualization functions to help users explore and understand complex data relationships. Through these innovations, AGENTiGraph aims to provide a more user - friendly, flexible, and efficient platform for complex knowledge management and analysis tasks.

AGENTiGraph: An Interactive Knowledge Graph Platform for LLM-based Chatbots Utilizing Private Data

GraphAgent: Agentic Graph Language Assistant

KG-Agent: An Efficient Autonomous Agent Framework for Complex Reasoning over Knowledge Graph

GraphTeam: Facilitating Large Language Model-based Graph Analysis via Multi-Agent Collaboration

Zero-shot Knowledge Graph Question Generation Via Multi-agent LLMs and Small Models Synthesis

Graphusion: A RAG Framework for Knowledge Graph Construction with a Global Perspective

Graphusion: Leveraging Large Language Models for Scientific Knowledge Graph Fusion and Construction in NLP Education

KGGPT: Empowering Robots with OpenAI's ChatGPT and Knowledge Graph.

Knowledge Graph Enhanced Language Agents for Recommendation

Chatlaw: A Multi-Agent Collaborative Legal Assistant with Knowledge Graph Enhanced Mixture-of-Experts Large Language Model

KnowAgent: Knowledge-Augmented Planning for LLM-Based Agents

GraphAide: Advanced Graph-Assisted Query and Reasoning System

BioKGBench: A Knowledge Graph Checking Benchmark of AI Agent for Biomedical Science

Enhancing Large Language Models with Knowledge Graphs for Robust Question Answering

Knowledge Solver: Teaching LLMs to Search for Domain Knowledge from Knowledge Graphs

Knowledge Graph and Deep Learning-based Text-to-GQL Model for Intelligent Medical Consultation Chatbot

Knowledge Graph Based Agent for Complex, Knowledge-Intensive QA in Medicine

An Interactive System for Knowledge Graph Search

Enhancing Dialogue Generation via Dynamic Graph Knowledge Aggregation

Generate-on-Graph: Treat LLM as both Agent and KG in Incomplete Knowledge Graph Question Answering