This workflow serves as a solid foundation when you need an AI Agent to return output in a specific JSON schema, without relying on the often-unreliable Structured Output Parser.
What It Does
The example workflow takes a simple input (like a food item) and expects a JSON-formatted output containing its nutritional values.
Why Use This Instead of Structured Output Parser?
The built-in Structured Output Parser node is known to be unreliable when working with AI Agents.
While the n8n documentation recommends using a “Basic LLM Chain” followed by a Structured Output Parser, this alternative workflow completely avoids using the Structured Output Parser node.
Instead, it implements a custom loop that manually validates the AI Agent's output.
This method has proven especially reliable with OpenAI's gpt-4.1 series (gpt-4.1, gpt-4.1-mini, gpt-4.1-nano), which tend to produce correctly structured JSON on the first try, as long as the System Prompt is well defined.
In this template, gpt-4.1-nano is set by default.
How It Works
Instead of using the Structured Output Parser, this workflow loops the AI Agent through a manual schema validation process:
- A custom schema check is performed after the AI Agent response.
- A runIndex counter tracks the number of retries.
- A Switch node:
- If the output does not match the expected schema, it routes back to the AI Agent with an updated prompt asking it to return the correct format. The process allows up to 4 retries to avoid infinite loops.
- If the output does match the schema, it continues to a Set node that serves as chat response (you can customize this part to fit your use case).
This approach ensures schema consistency, offers flexibility, and avoids the brittleness of the default parser.