Security Design and Evaluation of a Citi- zen-Oriented Generative AI Chatbot Using RAG (Retrieval-Augmented Generation)

抄録

This study develops and evaluates two Retrieval-Augmented Generation (RAG) chatbot systems (LangChain and Dify) for municipal services, offering deployment guidelines tailored to technical capacities. LangChain, a code-based framework, provides advanced customization for complex integration, whereas Dify is a GUI-based platform enabling rapid, low-code operation. Both systems were equipped with key security features—including PII detection, prompt injection prevention, and request rate limiting—and evaluated using simulated administrative inquiries. Our evaluation indicates that LangChain is the better choice for environments with strict security requirements and detailed management, while Dify is a practical option for municipalities with limited technical expertise due to its speed and ease of maintenance.