Transform OpenAI API Response into Clean JSON

Result: clean OpenAI response JSON

The example pipeline turns a nested OpenAI chat completion response into a smaller production-ready object.

Input:

{
  "id": "chatcmpl_9x7openai1049",
  "model": "gpt-4.1-mini",
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "{\"status\":\"paid\",\"summary\":\"Customer paid invoice INV-1049.\",\"next_action\":\"send_receipt\"}"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 817,
    "completion_tokens": 42,
    "total_tokens": 859
  }
}

Output:

{
  "model": "gpt-4.1-mini",
  "response_id": "chatcmpl_9x7openai1049",
  "message_text": "{\"status\":\"paid\",\"summary\":\"Customer paid invoice INV-1049.\",\"next_action\":\"send_receipt\"}",
  "finish_reason": "stop",
  "prompt_tokens": 817,
  "completion_tokens": 42,
  "total_tokens": 859
}

Open the matching OpenAI Response Cleaner template if you want to start from this workflow directly.

Why OpenAI responses need cleanup

An OpenAI API response is designed to describe the API transaction, not just the useful answer. That is why a single response can include:

• request identifiers
• model names
• one or more choices
• assistant messages
• finish reasons
• token usage
• system fingerprints
• timestamps

That metadata can be valuable, but most downstream systems need a smaller contract. A product database, analytics table, support log, or webhook step usually wants the answer text, the model, and the usage numbers.

Pipeline steps

This workflow uses two transformation steps:

1. Extract Response Fields: creates top-level fields for response_id, message_text, finish_reason, and token usage.
2. Keep Production Fields: removes the wrapper fields and keeps only the clean response object.

The result is easier to validate, diff, store, and review.

When to use this workflow

Use this workflow when you need to:

• extract content from an OpenAI response
• log token usage without storing the whole payload
• compare model outputs across runs
• send the answer into another automation step
• normalize OpenAI responses before analytics

If you are cleaning unpredictable model-generated JSON rather than the OpenAI API envelope, use the related LLM cleanup workflow next.

Why use a visual pipeline instead of a script?

A small script works when the shape never changes. A visual pipeline is better when you want the transformation to be inspectable, reusable, and safe for repeated API payloads.

ForgeJSON Pipeline lets you keep the cleanup steps visible:

• what gets extracted
• what gets removed
• what the final JSON looks like
• which fields downstream systems can rely on

That matters when AI responses are part of product workflows, not just one-off experiments.

FAQ

How do I extract content from an OpenAI API response?

The answer text is usually inside choices[0].message.content. This workflow extracts that content into message_text and keeps it next to the model, finish reason, and token usage.

Can I clean OpenAI responses without writing code?

Yes. Paste an OpenAI response into the pipeline, run the extraction steps, and copy the clean JSON output.

Does this parse JSON inside the message content?

This template extracts the message content as text. If the content itself contains JSON, you can add a follow-up cleanup or parsing step in ForgeJSON Pipeline.

What should I store from an OpenAI response?

For many product workflows, store the response id, model, message text, finish reason, and token counts. Keep the full raw response only when you need audit-level debugging.