Design and implementation of a role-playing dialogue agent for mobile phone after-sale service

Abstract

An intelligent after-sales dialogue agent has been designed to enhance both the service quality and operational efficiency. The system uses the ReAct (Reason + Act) cognitive model to integrate the reasoning ability of the large language model (LLM) with actual work. In this procedure, a constraint generation method based on function calls is applied, which can improve the precision of the relevant maintenance policies and service location data.

To deal with the redundancy and large amount of computation in multi-turn conversations, a dynamic dual-track memory management module (MemoryManager) has been set up, employing an information entropy algorithm for effective memory compression. Experimental data indicate that the agent can attain a tool-routing accuracy of 93.33%. The average total response time is maintained at 2.66 seconds, which satisfies the industrial standards. Additionally, the dynamic compression method can decrease the token usage by 60%-70% in long discussions in comparison with the conventional basic structures. By showing the tool invoking process, the system greatly improves the transparency and the users' trust.