DC-RAG: a dual-channel retrieval-augmented generation framework for audit analysis

With the continuous growth of information retrieval and knowledge acquisition demands, intelligent question-answering systems have been widely adopted across various vertical domains. However, in the context of audit result announcement analysis, domain-specific models tailored to the characteristics of audit texts remain scarce, rendering a large amount of structured and semantic information difficult to exploit effectively. To address this limitation, this paper proposes a dual-channel retrieval-augmented generation framework for audit result announcements, termed Dual-Channel Retrieval-Augmented Generation. The framework establishes a dual-path retrieval mechanism over a document database and a relational database, respectively retrieving candidate evidence from unstructured audit texts and structured knowledge graphs, and integrates multi-source evidence through a unified evaluation and re-ranking strategy, thereby enhancing the system's capability in audit semantic understanding and reasoning. Building upon this design, an end-to-end intelligent question-answering model is constructed by incorporating modules for query understanding, evidence re-ranking, and response generation, enabling knowledge retrieval and answer generation driven collaboratively by multiple data sources. Experimental results demonstrate that the proposed model consistently outperforms baseline methods in both retrieval effectiveness and answer quality, accurately interpreting user queries and producing high-quality responses, thus providing a feasible solution for the intelligent application of audit result announcements.