Entry, Exit, and Guard Nodes: Protecting the Graph

A graph is like a building. If you don't control the Entrance and the Exit, it's not secure. In LangGraph, we use specialized nodes to ensure that only valid data enters our agent's brain and only high-quality data leaves it.

1. The Entry Point: Input Sanitization

Don't let the user's raw prompt hit your expensive LLM brain immediately. Use an Entry Node to validate the input.

• Check: Is the prompt a "Jailbreak" attempt?
• Check: Is the prompt asking about a topic we are forbidden from discussing?
• Action: If the input is "Bad," route immediately to an Exit_Node with an error, bypassing the expensive AI reasoning.

---

2. Guard Nodes: The Internal Checkpoints

A Guard Node is a node that doesn't "Do" any work. It only "Checks" the work done by a previous node.

• Example: An Agent_Node writes some Python code. The Guard_Node then runs a linter or a security scanner on that code before it is actually executed.

---

3. The Exit Node: Final Verification

Before you show the answer to the user, the Exit Node performs one last check.

• Fact-Checking: Does the answer align with our RAG documents?
• Format Check: If the user asked for "ONLY JSON," does the final text contain only valid JSON?

---

4. Visualizing the Perimeter

graph LR
    START --> GuardIN[Entry Guard: Check for Safety]
    GuardIN -- Pass --> Brain[LLM Thinking]
    GuardIN -- Fail --> ERR[Exit: Forbidden Topic]
    Brain --> GuardOUT[Exit Guard: Check Factuality]
    GuardOUT -- Bad --> Brain[Refine Thought]
    GuardOUT -- Good --> END

---

5. Code Example: A Safety Guard Node

def input_guard(state: State):
    user_input = state["messages"][-1].content
    forbidden_words = ["hack", "illegal", "exploit"]
    
    if any(word in user_input.lower() for word in forbidden_words):
        return {"messages": ["Error: I cannot assist with that request."], "is_safe": False}
    
    return {"is_safe": True}

# In the graph configuration
workflow.add_conditional_edges(
    "input_guard",
    lambda x: "continue" if x["is_safe"] else "end",
    {"continue": "agent_node", "end": END}
)

---

Key Takeaways

• Entry Points are the first line of defense against bad prompts.
• Guard Nodes act as internal "Quality Gates" within the graph.
• Exit Verification prevents hallucinations from reaching the end user.
• Separating Safety into its own node makes the overall system more modular and easier to update.