How to erased the frist node's output if we have two nodes in a single chained-graph

[CharmYue]

We noticed that every time we received the second node's (Counseling_node) output, we found an unexpected output from the first node (classification node). We use this node to Rag and serve with LLM once, and then come out with a result, which is used as input for the second node. Could you help me with that? Thank you for your time and wisdom in advance. ：）

First and Second node are using the same LLM. We do not want the end user to see the first node's thinking process(output). We want direct output from the final node.

Here is the langgraph and the code.

[onestardao]

I’ve run into the same issue before — when chaining multiple nodes in a LangGraph, the intermediate node outputs end up visible even if we don’t want them shown. The fix is actually pretty straightforward:

In your classification_node, mark the output as internal or just control what gets passed explicitly:

return {"risk_desc": risk_desc, "_internal_only": True}

Then in your counseling_node, just clear out that field manually if you don’t want it lingering:

def counseling_node(state):
    # your logic...
    state.risk_desc = None  # wipe the intermediate output
    return {"final_plan": plan}

Or even more cleanly:

state.pop("risk_desc", None)

That removes the key entirely.

---

Quick tip:
LangGraph doesn’t automatically hide outputs between nodes — it assumes you want full traceability via AgentState. So if you’re preparing this for production or external users, you’ll want to explicitly return only the final fields at the last node.

If you ever need more advanced control (like field-level privacy across the graph), I’ve built a semantic firewall layer on top of LangGraph for exactly this. Happy to share more if that becomes a thing.

Good luck — hope your graph flows clean and your nodes stay quiet.

[CharmYue]

Thank you so much for your kind reply @onestardao , which is very helpful！

[onestardao]

Glad it helped!

Just in case you're building more complex LangGraph workflows — I recently compiled a structured map of common LLM issues (including graph traceability, memory drift, persona blending, etc.) and how to solve them at the reasoning layer. Might be useful if you plan to scale up or maintain reliability across sessions.

📎 ProblemMap – WFGY

Best of luck with your graph project — hope it flows clean and stays logically sharp!