GUIDE:企业级文档到工件生成的受控统一智能框架
文章背景与核心概要
企业准则文档通常具有高度的异构性和多模态特征,融合了叙述性文本、复杂表格及嵌入式图像。传统的各类大语言模型(LLM)和视觉语言模型(VLM)在处理此类输入时,往往面临幻觉内容、表格结构损坏以及缺乏从提取到验证及工件生成的全流程受控机制等挑战。目前,企业仍主要依赖人工处理,每份文档的处理周期长达2至3天。
为了解决这些痛点,本文提出了 GUIDE(Governed Unified Intelligence for Document-to-Artifact Generation,即“文档到工件生成的受控统一智能”)。这是一个基于共享版本化规则库的受控多智能体框架,通过模式验证的智能体间契约以及端到端的溯源追踪,实现了从文档解析到工件生成的高效、准确自动化,显著提升了企业处理复杂文档的效率与合规性。
执行摘要
企业准则文档通常是异构且多模态的,结合了叙述性文本、复杂表格和嵌入式图像。传统的大语言模型(LLM)和视觉语言模型(VLM)系统在处理这些输入时往往表现吃力,经常引入幻觉内容、导致表格结构退化,且缺乏涵盖从提取到验证再到工件生成的受控工作流。因此,企业一直依赖人工处理,每份文档需要耗费2到3天的时间。
Enterprise guideline documents are often heterogeneous and multimodal, combining narrative text, complex tables, and embedded images. Traditional Large Language Model (LLM) and Vision-Language Model (VLM) systems often struggle with these inputs, frequently introducing hallucinated content, degrading table structures, and lacking governed workflows that span beyond extraction to validation and artifact generation. Consequently, enterprises have relied on manual processing, consuming 2 to 3 days per document.
为了克服这些挑战,本文介绍了 GUIDE(Governed Unified Intelligence for Document-to-Artifact Generation),这是一个受控的多智能体框架,构建在共享的、版本化的规则库之上,具有模式验证的智能体间契约和端到端的溯源追踪功能。
To overcome these challenges, the paper introduces GUIDE (Governed Unified Intelligence for Document-to-Artifact Generation), a governed multi-agent framework built on a shared, versioned rule store with schema-validated inter-agent contracts and end-to-end provenance tracking.
核心亮点与架构
- 专业化多智能体框架: GUIDE 利用六个专业智能体来处理流水线的不同阶段:
- 解析
- 基于 VLM 的提取
- 一致性检查
- 评估
- 人机协同(HITL)升级
- 个性化定制的工件合成
- Specialized Multi-Agent Framework: GUIDE utilizes six specialized agents to handle distinct phases of the pipeline:
- Parsing
- VLM-driven extraction
- Consistency checking
- Evaluation
- Human-in-the-loop (HITL) escalation
- Persona-tailored artifact synthesis
- 严格的治理机制: 依赖于共享的版本化规则库、模式验证的智能体间契约以及全面的端到端溯源追踪,以消除内容幻觉和结构退化。
- Rigorous Governance: Relies on a shared versioned rule store, schema-validated inter-agent contracts, and comprehensive end-to-end provenance tracking to eliminate content hallucinations and structure degradation.
评估与结果
GUIDE 在 120 份真实企业准则文档上进行了评估,展现了强劲的性能指标: * 文档成功率: 在评估的文档中实现了 96% 的成功率。 * 规则提取: 提取了 3,896 条规则,其中 71.4% 实现了自动批准。 * 工件生产: 生成了 812 个可部署的工件。 * 效率: 将文档处理周期从几天大幅缩短至每份文档 40–125 分钟。
Evaluated across 120 real-world enterprise guideline documents, GUIDE demonstrated strong performance metrics: * Document Success Rate: Achieved a 96% success rate across evaluated documents. * Rule Extraction: Extracted 3,896 rules, with 71.4% automatically approved. * Artifact Production: Generated 812 deployment-ready artifacts. * Efficiency: Drastically reduced document turnaround time from days down to 40–125 minutes per document.
访问与资源
- 全文链接:
- 查看 PDF
- HTML 版本(实验性)
- TeX 源码
- 许可协议: 知识共享署名 4.0

- 引用与工具: 可通过 Google Scholar、Semantic Scholar 和 NASA ADS 获取。
- Full-Text Links:
- View PDF
- HTML Version (Experimental)
- TeX Source
- License: Creative Commons Attribution 4.0
- Citations & Tools: Available via Google Scholar, Semantic Scholar, and NASA ADS.