文章背景与核心概要
在构建面向目标的对话系统时,开发人员常常面临两难境地:既要准确回答事实性问题、捕获用户提供的信息,又要系统性地推进业务目标,同时还要避免对话陷入死板、不自然的问卷式交互。针对这一挑战,本文提出了一种名为“目标导向检索增强对话引擎”(GRACE)的全新架构,它有机结合了符号学、RAG(检索增强生成)与生成式AI,旨在实现可审计且高度自适应的对话编排。
GRACE 的核心在于其业务目标编译器(Business Goal Compiler),该编译器能够将原始的业务意图转化为不可变的业务目标、归一化优先级向量、规范化问题以及初始状态向量。在运行时,系统利用访客输入的证据动态更新状态完成度,并选择上下文调制的后续提示词,在满足最低访客效用约束的前提下,最大化预期的业务推进进展。实验表明,该架构在多语言、多领域的复杂对话场景中展现出了极高的准确性与鲁棒性。
面向目标导向对话编排的可审计符号-RAG-生成式AI架构
作者: Ramon Gonzalez, Antonio Diaz (Mentomy AI)
发表时间: 2026年9月7日
arXiv: 2609.07152 [cs.AI] | DOI: 10.48550/arXiv.2609.07152
📌 执行摘要
目标导向的对话系统往往难以在回答事实性问题、捕获用户提供的信息以及系统性推进业务目标之间取得平衡,同时容易演变成生硬、不自然的问卷调查形式。
Goal-oriented conversational systems often struggle to balance answering factual questions, capturing user-provided information, and systematically driving business objectives without devolving into rigid, unnatural questionnaires.
本文引入了目标导向检索增强对话引擎(GRACE),这是一种用于实现可审计且具适应性对话编排的新型符号-RAG-生成式AI架构。GRACE 使用业务目标编译器将业务意图翻译为不可变的目标、归一化优先级向量、规范化问题以及初始状态向量。在运行时,它利用访客编写的证据动态更新状态完成情况,并选择上下文调制的后续提示词,在遵守最低访客效用约束的同时,最大化预期的业务进展。
This paper introduces the Goal-oriented Retrieval-Augmented Conversation Engine (GRACE), a novel Symbolic-RAG-Generative AI architecture designed for auditable and adaptable conversation orchestration. GRACE uses a Business Goal Compiler to translate business intent into immutable objectives, normalized priority vectors, canonical questions, and initial state vectors. At runtime, it leverages visitor-authored evidence to dynamically update state completions and selects contextually modulated follow-up prompts, maximizing expected business progress while respecting a minimum visitor-utility constraint.
🧠 核心架构与方法论
- 业务目标编译器: 将原始业务意图转换为结构化的、受指令约束的框架,包含:
- 不可变目标集
- 归一化优先级向量
- 规范化问题
- 初始状态向量
- 运行时编排(GRACE): 评估完整的对话历史、最新的访客消息、当前状态,以及来自辅助检索增强生成(RAG)组件的扎实答案。
- 基于证据的状态更新: 仅使用访客编写的证据系统性地更新任务完成指标。
- 受约束的策略优化: 选择经过上下文调制的后续问题,以最大化预期的业务进展,同时满足严格的访客效用阈值。
- Business Goal Compiler: Transforms raw business intent into a structured, instruction-constrained framework comprising:
- Immutable objective sets
- Normalized priority vectors
- Canonical questions
- Initial state vectors
- Runtime Orchestration (GRACE): Evaluates the complete conversation history, latest visitor message, current state, and grounded answers from a secondary Retrieval-Augmented Generation (RAG) component.
- Evidence-Based State Updates: Systematically updates task completion metrics exclusively using visitor-authored evidence.
- Constrained Policy Optimization: Selects contextually modulated follow-up questions to maximize expected business progress while satisfying strict visitor-utility thresholds.
📊 评估与结果
作者在34场对话(总计119个协议定义的访客轮次)中对 GRACE 进行了评估,涵盖两个不同的领域: * 24场英语房地产对话 * 10场西班牙语专业清洁对话
The authors evaluated GRACE across 34 conversations totaling 119 protocol-defined visitor turns, encompassing two distinct domains: * 24 English real-estate conversations * 10 Spanish professional-cleaning conversations
关键性能指标
| 指标 | 性能得分 |
|---|---|
| 精确状态转换准确率 | 84.9% |
| 证据准确率 | 91.6% |
| 证据召回率 | 89.6% |
| 单调性 | 100% |
| 终端状态准确率 | 94.1% |
Key Performance Metrics
Metric Performance Score Exact State-Transition Accuracy 84.9% Evidence Precision 91.6% Evidence Recall 89.6% Monotonicity 100% Terminal-State Accuracy 94.1%
评估证明了其在标准工作流、多目标追踪、RAG绕路场景、验证检查、拒绝处理以及整体运营鲁棒性方面具有强大的符号状态性能。
The evaluation demonstrates robust symbolic-state performance across standard workflows, multi-goal tracking, RAG-detour scenarios, validation checks, refusal handling, and general operational robustness.
🔗 链接与资源
- 全文访问: 查看 PDF | HTML 版本
- 代码、数据与相关工具: 可通过 arXivLabs 集成门户 访问(包含 CatalystEx、HuggingFace 和 Connected Papers 工具)。
- Full-Text Access: View PDF | HTML Version
- Code, Data & Associated Tools: Accessible via the arXivLabs Integration Portal (including CatalystEx, HuggingFace, and Connected Papers tools).