第三代合作夥伴計畫（3rd Generation Partnership Project， 3GPP）標準問答的困難在於，答案所需證據常常只是龐大且持續演進的規範語料中的一小段內容。題目可能依賴特定Release 的定義、表格列、例外條件、協定欄位，或另一份規格中的交互參照。檢索式生成（Retrieval-Augmented Generation， RAG）適合作為此類標準文件的查詢介面，但檢索到主題相關段落並不代表檢索到足以區分選項的證據。當 RAG 系統答錯多選標準題時，單一總體準確率無法說明錯誤來自來源覆蓋不足、檢索錯誤、相關文件中遺漏關鍵子條款、作答模型誤讀證據，或 benchmark 標籤本身具爭議。 本論文提出證據充分性診斷框架及其來源覆蓋修復服務 SpecMend，以處理作答前未知或缺漏答案承載 TS/TR 的情況。只有可能的來源缺漏會啟動 SpecMend；它在不向來源路由器提供官方答案的情況下提出並取得官方 TS/TR、前處理、建立與正式專案索引隔離但在單次展示內共用的修復索引，並進行有界重答。官方 key 只供外部結果評分與狀態記錄。基礎語言模型保持固定，故本文的自我強化是改善資料庫與證據路徑，而非重新訓練模型。離線診斷結合語料存在性與詞彙檢查，以及在資訊洩漏控制下僅檢視檢索片段的大型語言模型評審。這些標籤是針對 3GPP 標準問答的診斷訊號，用以區分來源覆蓋缺口、檢索證據缺口、證據使用或grounding 失誤，以及 benchmark 或人工複核邊界。語意相似度基準則因可能將主題相關誤判為證據充分而被否決。 本論文以 TeleQnA 中標記為 Release 17/18 的本地完整 1，513 題 manifest 作為主要回顧式分析範圍，並在目前專案語料快照上進行評估。在含 292 份來源檔案、允許模型知識補充的 top-5 協定下，forced-choice 評分及其匹配的生成式重播皆達到 84.01%。在另一組匹配且僅依賴檢索證據的生成式作答 lineage 中，索引中的 TS/TR 來源檔案由 63 份擴充至 136 份，使準確率由 69.13% 提升至 81.82%；這是本專案內觀察到最大的匹配準確率變動，不是目前gate 的完整題集效益估計。另以同一組 30 題，分別從空索引及含 63 份來源檔案的既有索引啟動目前的來源取得適用性判定流程。這兩組展示提供機制執行與來源追溯證據。移除 Release標籤、匹配語料版本、改變檢索深度或證據粒度，皆使部分題目改善、另一些題目退步。因此，本論文不宣稱存在一項通用的後端策略。 整體而言，本論文將貢獻定位於固定語言模型周邊的 RAG 資料庫建構與來源修復，而非新的基礎模型或單一通用檢索策略。SpecMend 實作證據充分性診斷框架中的來源覆蓋修復分支，將錯題診斷轉化為候選規格提出、受控來源取得、前處理、隔離索引與有界重答。框架同時將檢索證據缺口、證據使用失誤及 benchmark 或複核邊界分開處理，但不將這些分支描述為同等完成的功能。主要 headline 與來源覆蓋結果屬回顧式且受語料快照影響；Release 及跨協定控制提供的是分離的 policy boundary，而非外部驗證。後續研究應在不提供 benchmark Release 標籤的較大型互斥題集上驗證 Release 選擇、比較由不確定性驅動的mixed 與 latest-version fallback、以迴歸檢查及來源提升或刪減規則管理新來源，並在具完整紀錄的互斥複核流程中檢驗診斷標籤的穩定性。;Question answering over 3rd Generation Partnership Project (3GPP) standards is difficult because the required evidence is often a small part of a large and changing specification corpus. A question may depend on a release-specific definition， a table row， an exception clause， or a protocol field in another document. Retrieval-augmented generation (RAG) is a useful interface for this task， but retrieving text about the right topic is not enough. The system must retrieve evidence that separates the correct option from the other options. When a RAG system answers a standards question incorrectly， accuracy alone cannot show whether the problem is a missing source file， a retrieval miss， a missing subclause， answerer misunderstanding， or an ambiguous benchmark item. This thesis proposes an evidence-sufficiency diagnostic framework and SpecMend， its implemented source-coverage repair service. The framework separates source-coverage gaps， retrieved-evidence gaps， failures in evidence use or grounding， and benchmark or review boundaries. Only a likely source-coverage gap activates SpecMend. For that branch， SpecMend proposes sources， acquires official specifications， preprocesses them， builds a repair index isolated from the canonical project index， and runs a bounded reevaluation; the index is shared across questions within each demonstration run. The source router does not receive the official answer; the benchmark key remains external and is used only for outcome scoring and evaluation-state bookkeeping. Other groups constrain actions but are not equally implemented repair branches. SpecMend builds or expands the database when needed Technical Specification or Technical Report (TS/TR) files are unknown or missing. The fixed base model is not retrained; self-improvement means strengthening its source and evidence path. Offline diagnosis combines source-presence checks， lexical anchors， and an LLM evidence judge that sees only retrieved chunks under leakage controls. Its labels are signals adapted to 3GPP QA rather than a universal RAG error taxonomy. A semantic-similarity baseline is rejected because topical similarity can look high even when the specification containing the needed evidence is absent. The main retrospective scope is the complete local TeleQnA manifest of 1，513 questions labeled Release 17/18， evaluated against the active project corpus snapshot. On the active 292-source snapshot under the fallback-permitting top-5 protocol， forced-choice scoring and its matched generation replay both reach 84.01%. In a separate matched retrieved-evidence- only generated-answer lineage， expanding the indexed TS/TR set from 63 to 136 source files raises accuracy from 69.13% to 81.82%. This is the largest matched in-project accuracy movement observed， not a full-set efficacy estimate for the current gate. Two demonstrations on the same selected 30-question set execute the current source-acquisition eligibility procedure from an empty index and from an existing index with 63 source files. They provide mechanism-execution and provenance evidence. Removing the release label， matching corpus versions， or changing retrieval depth or evidence granularity improves some questions and harms others. The thesis therefore does not claim one universal downstream policy. Overall， this thesis positions its contribution as RAG database construction and source repair for a fixed language model， rather than as a new base model or a universal retrieval policy. SpecMend implements the source-coverage repair branch of the diagnostic framework， converting wrong-case diagnosis into source proposal， guarded acquisition， preprocessing， isolated indexing， and bounded reevaluation. The framework keeps retrieved-evidence gaps， failures in evidence use， and benchmark or review boundaries separate， without presenting those branches as equally completed functions. The headline and source-coverage results are retrospective and snapshot-dependent; the release and cross-protocol controls define separate policy boundaries rather than external validation. Future work should validate release selection without benchmark release labels on larger disjoint sets， compare uncertainty-conditioned mixed and latest-version fallbacks， stage new sources with regression checks and promotion or pruning rules， and test diagnostic-label stability under documented disjoint review protocols.
