R&C: Design and Implementation of an Intelligent Document Question-Answering System with Integrated Multiple Training Models

With the evolution of AI technology and to support the ecosystem of domestic operating systems, this paper designs and implements an intelligent document question-answering robot system based on the Deepin domestic operating system to help users quickly and accurately obtain document information. By comparing different pre-trained language models, the system integrates RocketQA and ChatGLM models to enhance the understanding of Chinese queries and generate natural language answers. The system first uses the RocketQA model to vectorize the questions posed by users, combined with the Faiss library for preliminary retrieval and sorting. Then, it employs the RocketQA model for a second round of optimization on the retrieval results to further improve the relevance and accuracy of the answers. Finally, the system uses the ChatGLM model to transform the retrieved document content into fluent natural language answers, presenting them to users in the form of “reference link + answer”. Test results on the Deepin wiki and Linglong Usage dataset demonstrate that the system has the advantages of high accuracy and fast response speed, and provides a good user experience. The design concept is versatile and can be applied to related fields.