Abstract: To address the issues of a static and rigid knowledge base, weak semantic understanding, and poor integration of the question-answering module in the existing system of Wanjiazhai Water Conservancy Hub, a lightweight localized RAG-based dynamic knowledge base question-answering method is proposed. Based on 1,368 business documents from the Wanjiazhai and Longkou hubs, this method establishes a specialized water conservancy knowledge system covering four major domains,including company management and hub operation, with 16 subcategories. Using the Ollama and LangChain frameworks, the Qwen3-Embedding and Qwen3: 8B models are deployed locally to build a lightweight full pipeline of “text vectorization -semantic retrieval-augmented generation.” A multi-recall hybrid retrieval strategy is employed to combine the advantages of BM25 keyword matching and vector semantic retrieval, while a parent-child chunking strategy is adopted to preserve the hierarchical semantic structure of documents, and a context window management mechanism is introduced to enhance the coherence of multi-turn dialogues, thereby enabling full-lifecycle dynamic knowledge management and natural language intelligent question answering. The method requires no cloud dependency or high hardware investment, achieving continuous knowledge accumulation and on-demand reuse through pure localized lightweight deployment. RAGAS evaluation results show that the method achieves an average context precision of 91% and context recall of 89%, meeting the real-time consultation needs at engineering sites. This provides a replicable lightweight technical solution for the intelligent upgrade of knowledge services for similar water conservancy projects in the Yellow River basin.