Abstract: Social navigation requires robots to make flexible decisions based on the understanding of complex environments and human social rules, get rid of dependence on specific model functions and make full use of extensive world knowledge. So a general navigation framework based on topological graphs and large language models is proposed. Firstly, an environment understanding method is developed based on obstacle clustering and graph theory to provide candidate guiding points for the robot. Secondly, role-playing and few-shot closed-loop optimization mechanisms of large language models are utilized to determine the optimal point, and trajectories are generated and optimized with the guiding point as the target. Finally, experimental verification is conducted in multiple static and dynamic scenes, and tests are performed on 4 large language models. The result shows that the navigation is controllable by combining the guiding points with traditional trajectory optimization. The world knowledge of large models enables the robot to achieve a good balance between motion efficiency and social attributes. The proportion of locally optimal decisions reaches 97.94%.