RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Agents

Forum topic · 小凯 · 2026-09-20

Summary

RAFT is a stateful retrieval-augmented generation (RAG) framework for enterprise troubleshooting agents, addressing the limitation that conventional RAG treats support cases as static documents rather than multi-stage, stateful processes. The framework abstracts each closed historical case into a directed chain of timeline entries and retrieves at the entry level: when an entry matching the current case's intermediate state is found, RAFT returns the full parent-case trajectory anchored at that matched state. An optional case-level graph links cases through a configurable similarity representation. The authors evaluate the retrieval layer directly, avoiding costly end-to-end agent deployment. Due to scarce public multi-stage troubleshooting data, they build a benchmark synthesized from Microsoft Learn Windows Server documentation plus a real-world Apache Jira issue set with human-verified duplicate labels. RAFT outperforms vanilla RAG and GraphRAG baselines at every stage of case progression, with statistically significant gains over the strongest baseline; Jira results offer directional evidence of transferability to real case histories. Benchmarks, implementation, and the Jira evaluation set are publicly released. Paper: arXiv 2609.20754.

Overview

Field: Machine Learning
Authors: Mingxuan Zhang, Xiaowen Wang, Anupma Sharan, Zhengyi Chen, Chenyu Diana Zhang, Shanshan Yang, Chittibabu Pacharu
Published: 2026-09-17
arXiv: 2609.20754

Key Idea

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.

RAFT (Retrieval-Augmented Framework for Troubleshooting Agents) is a stateful RAG framework that:

Abstracts each closed historical case into a directed chain of timeline entries and retrieves at the entry level
Surfaces cases whose intermediate states match the active case
Returns the parent-case trajectory anchored at the matched state
Optionally links cases through a case-level graph with a configurable similarity representation

Evaluation

The retrieval layer is evaluated directly, without requiring production deployment of the full agent system. Because public multi-stage troubleshooting data is scarce, the authors construct:

1. A benchmark synthesized from Microsoft Learn Windows Server documentation 2. A real-world Apache Jira issue set with human-verified duplicate labels

RAFT outperforms vanilla RAG and GraphRAG baselines at every stage of case progression, with statistically significant gains over the strongest baseline. The Jira results provide directional evidence that the advantage transfers to real case histories.

The benchmarks, implementation, and Jira evaluation set are released publicly.

Original Abstract (excerpt)

> 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...

--- *Auto-collected on 2026-09-20*