ChatFlags: An AI-Powered Semaphore Interactive System

Abstract

This study presents the development of ChatFlags, an intelligent system for flag recognition and interaction. YOLO11 was selected as the visual backbone based on experiments involving five flag classification tasks. The custom dataset was refined and expanded to address the lack of publicly available resources. An improved model, YOLO-AKEMA, integrating attention mechanisms and adaptive convolution, achieved higher accuracy across 27 flag categories. The user interface was built by using the AI platform Dify, supporting conversational interaction. To mitigate hallucinations in large language models, a retrieval-augmented generation (RAG) framework was constructed by using curated flag documents and the BGE-M3 embedding model. Finally, the DeepSeek language model was integrated via workflow orchestration to complete the system. ChatFlags supports natural language dialogue, flag video analysis, knowledge quizzes, and text-to-image/video conversion. Its multimodal features enhance interactivity, offer a scalable solution for flag language education, and extend the integration potential of vision and language models.