現有混合檢索增強生成(Hybrid RAG)系統整合知識圖譜三元組與向量文本 時,常於檢索後處理階段引入資訊衝突與幻覺風險,根本原因在於缺乏對異構知 識的結構感知與系統性衝突處理。本研究提出「結構感知之診斷式混合式擴增檢 索後處理框架」(Diagnostic And Structure-Aware Post-Retrieval Framework, DASA), 以三個模組系統性解決:A1(異構知識校驗與評分)以 Cross-Encoder 結合 m× n 自然語言推理交叉驗證矩陣計算信心分數,並依閾值標記為衝突、互補或已確 認;A2(知識仲裁)依信心分數仲裁衝突知識群,整合出黃金事實集 G_golden; A3(骨架引導生成)先以大型語言模型提取邏輯骨架,再填充向量文本細節,並 以自我校對循環確保事實覆蓋率。 本研究以 MIAT 實驗室 244 篇碩士論文建構知識庫,自動生成 1,198 組 問答對(通過率 82.2%),分層抽樣建立 100 題評測子集,以 RAGAS 之忠實 度、答案相關性、脈絡精確率與召回率四項指標,搭配配對 t 檢定與 Wilcoxon 符號秩檢定評估統計顯著性。實驗結果顯示,DASA 在維持答案相關性與脈絡召 回率不退化的前提下,相較於 Hybrid-All 基線將忠實度提升 0.085(p < 0.001)、 脈絡精確率提升 0.143(p < 0.001)。更重要的是,此等增益可逐模組歸因:消融 實驗確認 A1 對脈絡精確率貢獻 +0.121(p < 0.01,Cohen′s d = 0.385)、A3 對 答案相關性貢獻 +0.108(p < 0.001,Cohen′s d = 0.352);A2 模組因採分階段驗 證策略之簡化實作,其完整分群仲裁設計列為優先未來工作。錯誤分析顯示 55% 問題 DASA 全面優於基線,26% 幻覺失敗源自本研究原創概念「覆蓋壓力」— —即骨架引導生成在高覆蓋率下的固有取捨。本研究之核心貢獻不在指標增益之 幅度,而在可解釋性與可控性:將 RAG 管線由黑盒推進至模組級可診斷性,使 各失敗模式可定位至特定設計決策點;且三模組可獨立啟閉、依任務與資源約束 調控組態,為知識密集型問答提供可診斷、可調控的異構知識整合框架。;Hybrid RAG combines knowledge graph (KG) triples with vector text to improve recall, but introduces post-retrieval conflicts that concatenation cannot resolve. This thesis proposes DASA (Diagnostic And Structure-Aware), a three-module post-retrieval framework: A1 scores each item via Cross-Encoder and an m×n NLI matrix (Confirmed/Complementary/Conflicting); A2 arbitrates conflicting clusters into a golden fact set (a stub; full arbitration is future work); A3 extracts a skeleton from hard facts and generates under it with a self-check loop. Evaluation uses 244 MIAT Laboratory theses (1,198 auto-generated QA pairs; 100-question stratified subset) on four RAGAS metrics—Faithfulness, Answer Relevancy, Context Precision, Context Recall—with paired t-tests and Wilcoxon signed-rank tests. Over the Hybrid-All baseline, DASA improves Faithfulness (+0.085) and Context Precision (+0.143, both p < 0.001) without degrading Answer Relevancy or Context Recall. Crucially, these gains are module-attributable: A1 drives Context Precision, A3 drives Answer Relevancy. The remaining 26% are hallucinations reflecting "Coverage Pressure," skeleton-guided generation′s inherent coverage–faithfulness trade-off. The primary contribution lies not in the magnitude of these gains but in interpretability and controllability: it turns post-retrieval processing from a black box into a module-level diagnosable, per-task configurable object of study, where each failure maps to a specific design decision.