生成后验证主导检索优化：RAG流水线特性的2^4因子消融
原标题：Post-Generation Verification Dominates Retrieval Optimization: A 2^4 Factorial Ablation of RAG Pipeline Features
AI 摘要
联合国大学研究者对RAG流水线的四个特性（章节扩展SE、智能体搜索AS、完整性检查CC、目录引导检索ToC）进行了2^4全因子消融实验，覆盖16种配置、24个查询、8种交互类型和2个云级模型，共768个条件。结果显示生成后验证（CC）在准确率、完整性和有用性上全面占优，单独使用CC（4.31/5）优于任何不含CC的配置，包括SE+AS+ToC三特性组合（4.11）。ToC以零LLM成本带来显著提升，AS效果小且不稳定，SE中性；最高质量配置延迟约为基线的两倍，形成六配置的帕累托前沿。作者结论是验证答案比优化检索更重要，且特性效用强烈依赖查询类型，需用多样化查询类型的因子设计来评估RAG特性。
正文节选
Post-Generation Verification Dominates Retrieval Optimization: A Factorial Ablation of RAG Pipeline Features Abstract Modern RAG pipelines stack many enhancement features, but these features are typically validated in isolation, leaving their interactions unmeasured. We run a full factorial ablation of four pipeline features—section expansion (SE), agentic search (AS), completeness check (CC), and table-of-contents-guided retrieval (ToC)—across 16 configurations, 24 queries spanning eight inter