Transformation LLM

Transforms data using an LLM prompt. Combines the power of a language model with structured data transformation.

Parameters

Parameter	Type	Required	Variable	Description
inputArray	dynamic value	Yes	Yes	Array of data to transform using the AI model.
itemVariable	text	No	No	Variable name representing each item in the prompt template. (Default: "item")
promptTemplate	text	Yes	Yes	Transformation instructions applied to each item. Supports {{item}}.
systemPrompt	text	No	Yes	System instructions sent before the transformation prompt.
model.provider	text	Yes	No	AI model provider (e.g. openai, anthropic).
model.model	text	No	No	Model identifier to use (e.g. gpt-4o-mini).
parameters.temperature	number	No	No	Model creativity (0 = deterministic, 1 = creative). (Default: 0.7, min 0, max 2)
parameters.maxTokens	number	No	No	Maximum number of tokens per individual response. (Default: 500, min 1, max 128000)
batchSize	number	No	No	Number of items processed per batch. (Default: 10, min 1, max 100)
parallelism	number	No	No	Number of batches processed in parallel. (Default: 3, min 1, max 10)
outputFormat	choice (text, json, structured)	No	No	Output format per item: plain text, JSON, or typed structure. (Default: "text")
outputField	text	No	No	Field name containing the LLM result in each item. (Default: "llmResult")
retryOnError	boolean	No	No	Automatically retry on model errors. (Default: true)
maxRetries	number	No	No	Maximum number of retries on error. (Default: 2, min 0, max 5)
outputVariable	text	No	No	Output variable name containing the transformed items.

Parameters marked Variable = Yes accept the {{blockName.field}} syntax.

Output

Output variable : llmTransformResult

{
  "success": false,
  "results": [],
  "totalItems": 0,
  "tokensInput": 0,
  "tokensOutput": 0,
  "totalTokens": 0,
  "durationMs": 0
}

Example

Normalize addresses with an LLM.

Input :

{"address": "12 rue victor hugo 75001 paris"}

Output :

{"content": "12 Rue Victor Hugo, 75001 Paris, France"}

Tip

Useful for semantic transformations impossible with code (spell correction, normalization, classification).