Skip to main content

​
How It Works

The AI Transformation node makes a direct call to a language model with a prompt you define. There is no agent persona or conversation history. The execution flow follows five steps:
  1. Resolves {{ }} templates in the input field and prompt with real execution data
  2. Assembles the model call with the prompt and configured parameters
  3. Waits for the response — 30-second timeout
  4. Parses the result: JSON when output_format is json, string when it is text
  5. Exposes the result via {{ $json.result }} for downstream nodes

​
Configuration Options

  • Input Field — {{ }} expression pointing to the text to process
  • Prompt — instruction for the model; accepts {{ }} templates with dynamic data
  • Model — gpt-4o-mini, gpt-4o, claude-3-haiku, or claude-3-sonnet
  • Output format — text for a free string or json for a structured object
  • JSON Schema — optional; guides the model on the expected fields when output_format is json
  • Temperature / Max Tokens — variation control (0.0 to 1.0) and token limit (100 to 4000)

​
Common Use Cases

  • Extract entities from unstructured forms or emails
  • Classify intent or sentiment to route service tickets
  • Summarize long tickets before notifying the team
  • Normalize date, phone, or address formats

​
Example

Prompt: 
Extract the name, email, and reason from the text below.
Return only valid JSON.

Text: {{ $trigger.body.mensagem }}
Output with output_format: json: 
{
  "result": {
    "nome": "Carlos Menezes",
    "email": "carlos@empresa.com",
    "motivo": "dúvida sobre faturamento"
  }
}

​
Tip

Use a temperature between 0.0 and 0.2 for extraction and classification — deterministic responses reduce failures in downstream nodes that consume the generated JSON.