用 LangGraph 构建客户旅程 Agent

有状态 Agent 图，跟踪客户阶段，自适应响应，需要时转人工。

工作流图

分步指南

1. 1

   Define the State Schema

   Create a TypedDict with fields: customer_id, stage (signup/onboarding/active), messages, customer_type, satisfaction_score. This state persists across all nodes in the graph.

2. 2

   Build the Graph Nodes

   Create nodes: greeting_node, assess_node, guide_node, qa_node, escalate_node. Each node receives the state, processes it, and returns updated state. Connect nodes with conditional edges.

3. 3

   Add Conditional Routing

   Use conditional edges: if satisfaction_score < 3, route to escalate_node. If customer asks a question, route to qa_node. Otherwise, continue to guide_node. This makes the agent adaptive.

4. 4

   Implement Memory Checkpointing

   Use LangGraph MemorySaver to persist conversation state. This allows the agent to resume conversations across sessions and maintain context. Store checkpoints in SQLite or Redis.

5. 5

   Deploy with FastAPI

   Wrap the LangGraph agent in a FastAPI endpoint. Add streaming for real-time responses. Deploy to your preferred platform (Vercel, Railway, self-hosted). Test with 20 simulated customer journeys.

标签

返回工作流教程库