🧠 Issue using Structured Output Parser with Cluster AI Agent and GPT-4o

[Yury_Maeda]

I’m using the Cluster AI Agent node with a Structured Output Parser to generate technical qualification questions for leads (based on a service name), and I’m getting a persistent parsing error when using more advanced or structured prompts with GPT-4o.

The same setup works fine with simpler prompts or when using GPT-3.5.

What is the error message (if any)?

rust

CopiarEditar

Model output doesn't fit required format

When I inspect the raw output, I see this:

json

CopiarEditar

{
  "response": {
    "generations": [
      [
        {
          "text": "",
          "generationInfo": {
            "finish_reason": "tool_calls"
          }
        }
      ]
    ]
  }
}

The "text" field is empty, and the model instead returns "finish_reason": "tool_calls", which breaks the Structured Output Parser that expects a JSON-formatted string in the text field.

Please share your workflow

This is the relevant part of the flow (simplified for clarity):

• Cluster AI Agent with GPT-4o
  • Prompt asks for 4 to 7 technical questions in JSON format
• Structured Output Parser with schema:

json

CopiarEditar

{
  "type": "object",
  "properties": {
    "perguntas": {
      "type": "array",
      "items": { "type": "string" }
    },
    "justificativa": { "type": "string" }
  },
  "required": ["perguntas", "justificativa"]
}

The workflow works perfectly with this prompt:

css

CopiarEditar

You are a technical analyst. Generate 4 to 7 clear, objective questions that help understand the scope of the following service. Explain briefly why they are relevant. Respond in the following JSON format: { "perguntas": [...], "justificativa": "..." }

But fails consistently when I use a more structured prompt (e.g., one using ###, markdown, or system-like instructions).

Share the output returned by the last node

json

CopiarEditar

{
  "response": {
    "generations": [
      [
        {
          "text": "",
          "generationInfo": {
            "finish_reason": "tool_calls"
          }
        }
      ]
    ]
  }
}

Expected output:

json

CopiarEditar

{
  "perguntas": [
    "What is the estimated area for service execution?",
    "Is there access to power on site?",
    "What is the expected deadline?"
  ],
  "justificativa": "These questions help qualify the lead's urgency, technical readiness, and project scope."
}

Information on your n8n setup

• n8n version: 1.76.3 (Self-Hosted)
• Database (default: SQLite): Default (SQLite)
• n8n EXECUTIONS_PROCESS setting (default: own, main): own
• Running n8n via (Docker, npm, n8n cloud, desktop app): Docker
• Operating system: Linux

---

How I fixed it (temporary workaround):
I used two GPT-based assistants:

• N8N Assistant (By Nskha) — helped analyze the issue and prepared a report for my prompt engineer.
• System Prompt Generator (by neural.love) — revised the system prompt to avoid triggering tool_calls.

After rewriting the prompt to remove markdown, formatting, and tool-style instructions — and explicitly asking for “plain JSON text with no function calls” — the issue was resolved.

[Shadoxity]

Thank you for this! I was having the same problem.

With 4o-mini it would work perfectly but 4o would not run and just error.

Getting rid of the structures output parser and simplifying the system message worked for me as well!

But I cannot seem to get 4o to give me a good consistent output like when using the structured output parser.
It just puts it in output instead of the json fields I want

[niwa8124]

Thanks for this thread. After running several trials, the structured output parser would almost always fail. The only times it didn’t fail were when the JSON was formatted precisely as the parser was defined. Any null sets caused the tool to fail.

I removed the tool and incorporated the JSON example into the system prompt. Turns out the AI agent doesn’t need to structured output parser tool to format output properly in my case.