Integration of LLM and Knowledge Graph: Building an Interpretable Intelligent Information Retrieval System

This article introduces an academic project that deeply integrates large language models (LLMs) with knowledge graphs. Through Retrieval-Augmented Generation (RAG) and graph reasoning technologies, it addresses the issues of LLM hallucinations, limitations in structured knowledge understanding, and insufficient interpretability, achieving a more accurate and interpretable intelligent information retrieval system.

Current LLMs perform well in open-domain problem handling, but they have pain points such as difficulty in ensuring factual accuracy, lack of explicit reasoning ability for structured knowledge, and insufficient interpretability of outputs. Traditional vector retrieval-based RAG alleviates some problems but is limited by semantic matching accuracy. Knowledge graphs store knowledge in the form of triples and have the characteristics of precision, interpretability, and reasoning ability. The integration of the two is an important direction to improve the reliability of AI systems.

The system adopts an end-to-end architecture, with processes including document loading, intelligent chunking, LLM triple extraction, context proximity analysis, graph construction and merging, community detection, and interactive visualization. Tech stack: Mistral-7B (locally deployed via Ollama) + LangChain framework; document chunking using RecursiveCharacterTextSplitter; triple extraction in JSON format; context proximity analysis to capture implicit co-occurrence relationships; NetworkX for graph construction, Girvan-Newman for community detection; PyVis for interactive visualization; CSV for caching intermediate results; Jupyter Notebook as the development environment.

The system is suitable for scenarios such as medical literature analysis and legal document review. It can identify core concepts, reveal hidden relationships, separate topic clusters, and support knowledge exploration. Compared to pure vector RAG systems, its advantages lie in the interpretability of outputs and structured reasoning ability—users can see the reasoning path of the answers.

This project demonstrates the technical path of integrating LLMs with knowledge graphs and provides a feasible engineering solution to solve the problem of large model hallucinations. Through structured knowledge representation and graph reasoning, it significantly improves output accuracy and interpretability while maintaining language understanding capabilities, which has important reference value for the construction of enterprise-level knowledge bases and intelligent question-answering systems.

The current implementation is limited to concept-level knowledge extraction and has limited ability to handle complex events and temporal relationships. Future explorations can include: introducing temporal knowledge graphs to support dynamic updates, combining vector retrieval to implement a hybrid RAG architecture, and developing question-answering systems that support multi-hop reasoning.