STEM与日常科技·英语精读30篇（5）
22 / 30
正在确认阅读权限…
RAG Failures in Customer Service Chatbots: When Retrieval Misleads Generation
客服聊天机器人中的RAG失效：检索环节如何误导生成结果
-
Retrieval-Augmented Generation systems retrieve policy documents before generating replies, yet outdated PDFs often rank higher than revised web content due to legacy metadata.检索增强生成系统在生成回复前会检索政策文档，但因遗留元数据，过时的PDF文件排名常高于修订后的网页内容。
-
A query about 'return window extension' may retrieve a discontinued holiday policy instead of current standard terms, causing misleading confidence in the response.关于‘退货期限延长’的查询可能检索到已停用的节假日政策，而非当前标准条款，导致回复可信度误导性偏高。
-
Semantic similarity scoring fails when product names change—'Galaxy S23 FE' and 'Galaxy S23 Fan Edition' yield low match scores despite identical coverage.当产品名称变更时，语义相似度评分失效——‘Galaxy S23 FE’与‘Galaxy S23 Fan Edition’实际保障范围相同，却得分极低。
-
Users reporting billing issues frequently trigger irrelevant warranty clauses because retrieval engines over-index on financial keywords like 'charge' or 'refund'.用户报告账单问题时，检索引擎常因过度匹配‘收费’‘退款’等财务关键词，错误触发无关的保修条款。
-
Post-hoc audit logs reveal that 68% of incorrect RAG outputs stem from retrieval top-k contamination—not generation hallucination.事后审计日志显示，68%的RAG错误输出源于检索Top-K结果污染，而非生成阶段的幻觉。
-
Effective mitigation combines timestamp-aware ranking, domain-specific query rewriting, and explicit uncertainty signaling in low-confidence retrieval contexts.有效缓解策略需结合时间戳感知排序、领域定制化查询重写，以及在低置信度检索场景中显式提示不确定性。