DianGPT: Bridging Precision and Efficiency for a Domain-Specific Question-Answering System

In this paper, we present DianGPT, a large language model (LLM) tailored for specialized question-answering tasks through integration with organizational archives and domain-specific knowledge bases of the Dian Group. To address the limitations of existing models in balancing precision with resource efficiency, the proposed system combines parameter-efficient Low-Rank Adaptation (LoRA), Retrieval-Augmented Generation (RAG), and structured data analysis. Our approach ensures the model dynamically retrieves contextual information while maintaining computational efficiency. Our experimental results demonstrate that DianGPT, when augmented with RAG, achieves marked improvements in response quality and task-specific performance, validated through rigorous benchmarking. The proposed framework also emphasizes evaluation transparency, prioritizing traceable metrics to assess reliability. By advancing methodologies for domain-specific language models, we explore practical engineering techniques for developing intelligent Q&A systems in specialized fields while offering scalable insights for adapting LLM in diverse applications. As a case study, DianGPT demonstrates the potential of hybrid fine-tuning and retrieval strategies to balance performance with resource efficiency in real-world deployments.