Constructing a Shopping Mall Customer Service Center Robot Based on the LLAMA-7B Language Model

Abstract

This study develops a customer service robot powered by a large language model, deployed at Tainan Spinning Mall. The system runs on the Temi robot and leverages Llama3 fine-tuned using PEFT methods like LoRA to optimize performance on an NVIDIA L4 GPU with limited memory. Retrieval-Augmented Generation (RAG) enhances response accuracy, while training data includes mall-specific datasets, such as store directories, product listings, and promotions, alongside general dialogue data. After model adaptation, it achieves 95% accuracy in answering mall-related inquiries, responding within five seconds, ensuring efficient and accurate assistance in fast-paced retail environments.