Skip to main content
A transformation changes a workflow’s data after extraction and before delivery. Use one to translate a field into English, summarize a long description, classify records by your own rules, or compute a field from other fields.

​
How a transformation works

A transformation has one or more steps. Steps run in order after extraction, and each step reads the output of the step before it. A step is one of two kinds:
  • AI instructions: a model fills each output field by following a written instruction, for example Translate the title into English. The step can read the whole record or only the fields you name as input.
  • Code or SQL: fixed logic that gives the same result every run, for example splitting a full name into first and last name.
Fields the transformation does not output pass through unchanged. Data quality checks run on the transformed data.

​
Functions and transformations

Both add or change fields. They differ in where the logic comes from. Start with a function when one fits. Use a transformation when no function does what you need.

​
Add or change a transformation

Ask the agent in the workflow, for example Add a titleEnglish field with the title translated into English. The agent writes the step and tests it on your data. If your workflow has no agent, ask in the Support Center.

​
Review a transformation

Open the Pipeline panel and select the Transform tab. Computed fields lists each field the transformation outputs, with a Code or AI badge. Select a field to see what the step does and the fields it is made from. Through the API, Get workflow by ID returns the transformation object:
  • fields: each output field with its name, type, summary, step, and inputs. The summary is the field’s instruction; fields from code or SQL steps have none. step is the type of step that produces the field, llm_transform or code. inputs lists the fields the value is made from, and is empty when the step reads the whole record.
  • contextFields: the fields the AI steps read as input.

​
Limits

  • You cannot edit a transformation directly in the dashboard or through the API.
  • Realtime monitors do not run transformations.