Abstract:The large language model (LLM) demonstrates excellent capabilities in natural language understanding and generation. However, it still faces challenges such as insufficient factual accuracy, difficulties in knowledge updating, and a lack of high-quality domain-specific datasets in knowledge-intensive tasks. To address these challenges, retrieval-augmented generation (RAG) has emerged as an effective solution. However, when applied to knowledge-intensive tasks in the carbon domain, RAG technology has limitations, including potential bias in query understanding, rigid external knowledge retrieval strategies, poor correlation between retrieved results and actual needs, and a lack of specific datasets for evaluating question-answering performance. To tackle these issues, this study proposes a Multi-pipeline-based RAG method, which utilizes the graph-enhanced recursive intelligent merge retrieval method to effectively improve retrieval accuracy. For the lack of Q&A datasets in specific domains, a large model-based approach is proposed to automatically generate Q&A datasets from the parent node text. Moreover, this study evaluates the following aspects using the text understanding capability of LMM, alongside traditional evaluation metrics such as precision and recall: (1) response-context-query correlation; (2) response-query correlation; (3) context-query correlation, and (4) loyalty evaluation. Experimental results show that the Multi-pipeline-based RAG method based on the GLM-4-Plus model achieves an accuracy of 85%, outperforming BM25-based RAG, Vector-based RAG, and Recursive-based RAG methods.