Sample LangGraph Support Agent

A multi-node customer support chatbot built with LangGraph. Serves as a canonical example of LangGraph features and as a test fixture for the AI Workflow Visualizer.

Workflow

START → [Guardrail] → [Router] ──billing──→ [Billing Agent] ──→ [Response Formatter] → END
                               ──technical→ [Tech Agent]    ──→ [Response Formatter] → END
                               ──complaint→ [Complaint Agent]──→ [Response Formatter] → END
                               ──escalate→ [Human Node (interrupt)] → [Response Formatter] → END

The Guardrail node rejects harmful or off-topic input before anything else runs. Each agent uses one or more @tool-decorated functions in a ReAct loop (agent → tools → agent) before handing off to the formatter.

Features Demonstrated

LangGraph Feature	Location
StateGraph + TypedDict state	src/state.py
add_node / add_edge	src/graph.py
add_conditional_edges	src/graph.py
START / END constants	src/graph.py
@tool decorated functions	src/tools/
bind_tools on ChatOpenAI	src/nodes/*_agent.py
ToolNode (ReAct loops)	src/graph.py
MemorySaver checkpointer	src/graph.py
interrupt() human-in-the-loop	src/nodes/human_node.py
Guardrail pattern	src/nodes/guardrail.py

Quickstart

# 1. Clone and install
pip install -r requirements.txt

# 2. Set credentials
cp .env.example .env
# Fill in OPENAI_API_KEY and TAVILY_API_KEY

# 3. Run a query
python - <<'EOF'
from dotenv import load_dotenv
load_dotenv()

from langchain_core.messages import HumanMessage
from src.graph import graph

config = {"configurable": {"thread_id": "demo-1"}}
result = graph.invoke(
    {"messages": [HumanMessage(content="Why was I charged twice this month?")]},
    config=config,
)
print(result["messages"][-1].content)
EOF

Human-in-the-Loop Escalation

When the router classifies intent as escalate, the graph pauses at human_node:

from langgraph.types import Command

# First call — graph pauses
graph.invoke({"messages": [HumanMessage(content="I need to speak to a human")]}, config=config)

# Human provides input — graph resumes
graph.invoke(Command(resume="I understand your frustration. Let me help you directly."), config=config)

Tests

pytest tests/ -v

Project Structure

src/
├── graph.py          # StateGraph assembly — the file the visualizer parses
├── state.py          # SupportState TypedDict
├── nodes/            # One file per graph node
└── tools/            # @tool-decorated functions
tests/
notebooks/
└── demo.ipynb        # Interactive walkthrough

---

Built by Trango Compute