[论文] RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Age...

论文概要

研究领域: ML 作者: Mingxuan Zhang, Xiaowen Wang, Anupma Sharan, Zhengyi Chen, Chenyu Diana Zhang, Shanshan Yang, Chittibabu Pacharu 发布时间: 2026-09-17 arXiv: 2609.20754

中文摘要

企业客户支持中的故障排查 Agent，其有效性依赖于从历史相似案例中检索可操作的指导，但现有检索增强生成（RAG）系统把支持案例当作静态文档，忽视了其多阶段、有状态的本质。我们提出 RAFT——面向故障排查 Agent 的有状态 RAG 框架：把每个已关闭历史案例抽象为时间线条目的有向链，在条目级别检索；当命中与当前案例中间状态匹配的条目时，返回以该状态为锚点的父案例完整轨迹；可选的案例级图通过可配置的相似度表示把案例彼此连接。我们直接评估该检索层——与评估整个 Agent 系统不同，这无需生产部署。由于公开的多阶段故障排查数据极为稀缺，我们构建了由 Microsoft Learn Windows Server 文档合成的基准，以及带有人工重复标记的真实 Apache Jira 问题集。RAFT 在案例推进的各个阶段均优于普通 RAG 和 GraphRAG 基线，且相对最强基线的增益统计显著；Jira 结果提供了优势可迁移到真实案例历史的方向性证据。我们发布了基准、实现与 Jira 评估集。

原文摘要

Effective troubleshooting agents in enterprise customer support depend on retrieving actionable guidance from similar historical cases, yet existing retrieval-augmented generation (RAG) systems treat support cases as static documents and overlook their multi-stage, stateful nature. We introduce RAFT (Retrieval-Augmented Framework for Troubleshooting Agents), a stateful RAG framework that abstracts each closed historical case into a directed chain of timeline entries and retrieves at the entry level, surfacing cases whose intermediate states match the active case and returning the parent-case trajectory anchored at the matched state; an optional case-level graph links cases through a configurable similarity representation. We evaluate this retrieval layer directly, which, unlike evaluating ...

#论文 #arXiv #ML #小凯