关键词: 大语言模型（LLM）, 知识问答系统, 大模型幻觉, 信息检索, 提示学习

Abstract: As global climate change becomes increasingly severe, enterprise carbon emission analysis has become a focal point of international attention. Addressing the issues of outdated knowledge updates in general-purpose large language models (LLM), the lack of professionalism and accuracy in augmented generation architectures when handling complex problems, and the high hallucination rate in the outputs of large-scale models, this paper proposes an enterprise carbon emission analysis and knowledge question-answering (Q&A) system based on large language models, by constructing a proprietary knowledge base. The paper introduces a diversified indexing module construction method to build a high-quality knowledge and regulatory retrieval dataset. For the knowledge Q&A task in the domain of carbon emission reports (policies), the paper proposes a self-prompting retrieval-augmented generation architecture, integrating intent recognition, improved structured chain-of-thought, hybrid retrieval techniques, high-quality prompt engineering, and Text2SQL system. This supports multi-dimensional analysis of enterprise sustainability reports, providing an efficient and accurate knowledge Q&A solution for carbon emission reports (policies). Through the multi-layer chunking mechanism, document indexing, and hallucination detection functionality, the paper ensures the accuracy and verifiability of the results, reducing the hallucination rate of LLM technology in this system. Through comparative experiments, the algorithm proposed in this paper demonstrates superior performance across various metrics in retrieval-augmented generation experiments under the synergy of all modules. It exhibits significant advantages in key information extraction and report evaluation for enterprise carbon emission reports, particularly in handling long-text.

Key words: large language model (LLM), knowledge question-answering system, hallucination, information retrieval, prompt learning