The analysis methods of big data and intelligence have improved the ability to mine and distinguish risks in power customer service. However, the continuous enrichment of service channels and the rapid updating of interactive content have led to low accuracy in identifying user demands in question and answer services, as well as insufficient ability to analyze user intentions, resulting in frequent customer service risks. In response to this issue, this paper proposes a method for constructing a large language professional model suitable for risk identification in power customer service, based on language model+RAG technology and fine-tuning model parameters to improve the accuracy of customer service risk identification. By inputting multi-modal data of power customer service, extracting semantic features of power customer service risks, developing a professional prompt word pattern framework for power customer service, concatenating vectored semantic features, and injecting risk semantic knowledge based on P-Uni. In the data correction stage, additional feature parameters are added to the adapter layer to fine tune the model, and the fine tuning parameters are selected and optimized based on the degree of tilt of the recognition effect, ultimately finding the optimal model configuration parameter group. Electric customer service involves a wide range of professional knowledge, and the factors affecting risk identification in customer service are complex and variable. Traditional language models need to be continuously trained and enhanced to improve the training effectiveness and risk identification accuracy of samples. This method reduces the training time of the model through a lightweight parameter tuning method, which greatly reduces the difficulty of knowledge learning and can provide a method reference for the application of risk management in power customer service.