1.4 KiB
1.4 KiB
title, type, tags, sources
| title | type | tags | sources |
|---|---|---|---|
| Intent Classification | concept |
Intent Classification
Definition
AI 自动识别客户消息的意图（Purpose / Goal），将非结构化自然语言映射为预定义的分类标签，从而触发对应的处理策略和回复逻辑。
Common Intent Categories
| Intent | Trigger Keywords | Handling Strategy |
|---|---|---|
| FAQ | "how much", "do you offer", "what is" | 从知识库检索答案并回复 |
| Appointment | "book", "schedule", "available", "reservation" | 检查可用性并确认预约 |
| Complaint | "frustrated", "refund", "terrible", "problem" | 标记人工审核 + 确认收到 |
| Review | "google review", "rating", "stars" | 感谢反馈 + 回应问题 |
| Order Status | "where is my order", "tracking", "delivery" | 查询系统返回状态 |
Implementation
Intent Classification 通常通过 LLM 的 Function Calling 或 Prompt Engineering 实现：
User Message → LLM Intent Classification → Match Intent to Handler → Generate Response
在 OpenClaw 中通过 AGENTS.md 配置 Classify intent 步骤实现。
Related Concepts
- Unified-Inbox：意图分类应用于统一收件箱中的每条消息
- Business-Knowledge-Base：FAQ 意图依赖知识库检索
- Human-Handoff：投诉意图触发人工升级