Abstract
This paper proposes CD-RAG (Chain-of-Thought and Dynamic-completion enhanced Retrieval-Augmented Generation), a graph-enhanced generative question answering system tailored for component-level process knowledge in industrial settings. Designed to improve semantic understanding and knowledge coverage in complex domain-specific tasks, CD-RAG integrates structured graph retrieval with large language model (LLM)-based generation into a hybrid QA framework. To address performance bottlenecks in conventional RAG systems under conditions of sparse entity-relation representation and incomplete knowledge graphs, CD-RAG introduces two key innovations: (1) a reasoning module guided by chain-of-thought prompting, which enhances multi-hop inference and causal response generation; and (2) a dynamic knowledge completion mechanism, which automatically triggers external web-based search upon retrieval failure, enabling knowledge re-ingestion and semantic closure. The system constructs a multi-dimensional knowledge graph from heterogeneous process-related texts, including technical manuals and industrial e-books, capturing information on workflows, operational standards, and part attributes. Experimental results demonstrate that CD-RAG consistently outperforms baseline systems across all evaluated dimensions. This work offers a generalizable framework and engineering pathway for deploying generative QA systems in real-world industrial knowledge environments.
Showing the abstract — retrieve the full paper via the Exa API.