An Intelligent Knowledge Base Construction Method Integrating a Vector Database and a Large Language Model

With the rapid proliferation of organizational documents, technical manuals, question-and-answer records, and regulatory materials, knowledge base systems are increasingly required to retrieve relevant information efficiently and provide a reliable evidential foundation for generated responses. Existing retrieval-augmented generation methods, however, continue to exhibit notable deficiencies in evidence organization and knowledge synchronization, frequently resulting in incomplete evidence coverage, insufficient answer grounding, or outdated retrieval outcomes. To address these limitations, this study proposes the VEC-KB framework, an intelligent knowledge base construction approach that integrates a vector database with a large language model. The framework introduces several key innovations: a semantic-structure-aware document segmentation strategy that preserves contextual integrity, a query-adaptive re-ranking mechanism that enhances retrieval precision, an evidence evaluation module that strengthens answer faithfulness, and a version-aware incremental synchronization protocol that ensures timely and efficient knowledge updates. Experimental evaluations conducted on the HotpotQA and Natural Questions datasets demonstrate that VEC-KB consistently outperforms established benchmark methods in overall performance, with particularly pronounced improvements in answer quality and evidence credibility, while maintaining robust recall capabilities. Furthermore, under perturbed conditions, VEC-KB exhibits significantly smaller performance degradation compared to competing approaches, confirming its superior stability. Collectively, the proposed framework achieves an effective balance among response quality, evidence reliability, and system robustness, offering a comprehensive and scalable solution for intelligent knowledge base construction.