云原生 Graph-RAG 错误归因解耦：数据完整性诊断框架

AI 摘要

该论文提出一个三层解耦诊断框架，用于将云原生 Graph-RAG 系统的错误正交归因于推理损失、知识图谱缺陷和 Cypher 生成错误。作者在青藏高原东南部时空生态知识图谱上注入八类缺陷进行评测，发现数据完整性而非算法推理是主要性能瓶颈，结构缺陷使系统准确率从 0.93 降至 0.39。研究还观察到参数化知识掩蔽效应（PKME），即 LLM 用内部记忆补偿断裂的检索路径，使表面查询生成错误减少逾 70%，掩盖了实际存储退化。

正文节选

Decoupling Error Attribution in Cloud-Native Graph-RAG: A Data Integrity Diagnostic Framework Thanks: This research was funded by Science and Technology Projects of Xizang Autonomous Region, China, project title ”An Intelligent Question-Answering System for the Ecological Environment of the Qinghai-Tibet Plateau Based on Large Language Model-Powered Knowledge Graph Agents”, grant number XZ202502ZY0073. Abstract Graph-RAG systems often assume pristine data quality, overlooking the severe impact o

发布时间：2026-09-15 12:00

抓取时间：2026-09-15 12:18

来源机构：arXiv