합성 지식그래프 기반 GraphRAG 기법을 응용한 민감정보 Q&A 챗봇

A Chatbot for Q&A of Sensitive Information Applying the Composite Knowledge Graph-based GraphRAG

In addition to the positive evaluations of LLMs' language understanding and generation capabilities, concerns have been raised about the potential leakage of sensitive information through questions entered into prompts. To address this, a method has been proposed that stores sensitive information in a separate knowledge graph to prevent LLMs from accessing it. However, storing various types of sensitive information in a single knowledge graph presents practical limitations. To account for the diversity and volume of sensitive information, multiple knowledge graphs need to be federated to construct correct answers. This study presents a GraphRAG-based approach by implementing a prototype chatbot using a composite knowledge graph, which consolidates multiple knowledge graphs that store different types of sensitive information to generate responses to input questions. The implemented chatbot converts natural language questions into query statements suitable for the knowledge graph via the LLM, which then executes queries on the knowledge graph to retrieve results. The LLM is solely involved in generating queries appropriate for the knowledge graph, while the chatbot handles the query execution and result extraction, thereby eliminating the possibility of sensitive information being introduced into the LLM. To validate the proposed method, the performance of the implemented chatbot was tested, confirming not only that correct responses were generated but also that the LLM effectively learned the content and structure of the composite knowledge graph using the approach developed in this study.