本研究結合 LLM-as-a-Judge、RAG 上下文指標、程式化檢索指標與人工檢測進行評估。結果顯示，低資源 SFT 與 Naive RAG 並不必然帶來增益；CPT 較能提升法律語境適應能力，而 CPT+SFT 在無 RAG 條件下未必優於 CPT，但結合 RAG 後可改善人工檢測表現。RAG 消融結果顯示，查詢轉換與重排序較能穩定提升檢索品質，RSE 主要提供上下文整理，代理式與知識圖譜輔助方法之效果則取決於法律爭點辨識、查詢規劃與證據組織能力。整體而言，訓練與 RAG 並非單純相加，其效益取決於語料品質、檢索精準度及模型整合外部證據的能力。本研究據此建立一套低資源中文法律問答的分層比較與評估框架。;Large language models can generate fluent text， but legal question answering additionally requires accurate legal concepts， statutory citations， constituent elements， and conclusions. For low-resource Traditional Chinese legal tasks， improving the reliability of open-source models under limited data and computational resources remains challenging.
This study investigates general-concept question answering in Taiwan criminal law through internal model adaptation and external legal knowledge augmentation. Experiment 1 compares Base， SFT， Base+Naive RAG， and SFT+Naive RAG. Experiment 2 evaluates hybrid retrieval， reranking， query transformation， Relevant Segment Extraction (RSE)， agentic RAG， and knowledge-graph-assisted structured hints. Experiment 3 compares Base， continued pre-training (CPT)， CPT+SFT， and their RAG variants to examine how training affects independent answering and the use of retrieved legal evidence.
Evaluation combines LLM-as-a-Judge， RAG context metrics， programmatic retrieval metrics， and manual legal assessment. Results show that low-resource SFT and Naive RAG do not necessarily improve performance. CPT more consistently enhances legal-domain adaptation， while CPT+SFT does not always outperform CPT without RAG but performs better in manual assessment when combined with RAG. Query transformation and reranking provide relatively stable retrieval gains， whereas RSE mainly improves context organization. Agentic and knowledge-graph-assisted methods depend strongly on legal-issue identification， query planning， and evidence organization. Overall， training and RAG are not simply additive; their effectiveness depends on corpus quality， retrieval precision， and the model′s ability to integrate external legal evidence.
