近年來，檢索增強生成（Retrieval-Augmented Generation， RAG）技術因能有效提升大型語言模型（Large Language Models， LLMs）的表現而受到廣泛關注。尤其在多跳問答（Multi-hop Question Answering）任務中，語言模型需要檢索系統能夠連結分散於不同文件及知識社群中的相關證據。然而，傳統的密集檢索式 RAG 雖能取得語意相近的文本，卻無法明確建構跨文件的證據推理鏈；另一方面，圖式傳播方法則容易受限於局部連結密集的社群，使檢索範圍侷限於特定區域。為了解決上述問題，本研究提出一種新穎的圖式檢索增強生成框架 SeriesRAG，將非結構化文本轉換為由實體、關係與文本片段所組成的加權知識圖。透過社群偵測與橋接節點選取機制，SeriesRAG 能有效識別跨社群的重要連結節點，同時降低圖傳播過程對高殘留區域的過度依賴。此外，本研究設計了 Dynamic QA Selection 機制，可從較低排序的候選文本中動態保留具代表性的錨點證據與互補證據，以提升檢索結果的完整性。實驗結果顯示，SeriesRAG 在多個真實資料集上的表現具有競爭力，並在多項常用評估指標上與現有最先進方法相比展現優異效能。實務上，SeriesRAG 可為企業知識搜尋、風險分析及法規遵循審查等應用提供兼具可解釋性與跨文件證據追溯能力的檢索架構。;Recently， Retrieval-Augmented Generation (RAG) techniques have attracted tremendous attention owing to the successful enhancement for the performance of LLM. Especially for multi-hop question answering， LLM model requires retrieval systems to connect evidence distributed across documents and knowledge communities. Conventional dense RAG retrieves semantically similar passages but does not explicitly model evidence chains， whereas graph propagation can become trapped in dense local communities. In this study， a novel graph-based RAG framework， SeriesRAG， is proposed to convert unstructured corpora into a weighted graph of entities， relations， and passages. We utilize community detection and bridge node selection to effectively expose cross-community connectors， while downgrade the propagation to high-residual graph regions. SeriesRAG also equips a dynamic QA Selection to retain active anchor and complementary passages from lower-ranked candidates. Experiments on various real datasets demonstrate the competitive performance comparing state-of-the-art baselines on several common evaluations. In practice， SeriesRAG could offer an explainable architecture for enterprise search， risk analysis， and compliance review for traceable cross-document evidence.
