Docs / Workflows / Actions / LLM call
LLM call
Run one model request to summarise, extract or rewrite the context supplied to a job.
Input and context
The prompt describes the result to produce. It can include event values, such as {{ properties.body }}. With Event context set to Full event, the model also receives the incoming event. Tagged properties only leaves that event out and supplies the values referenced in the prompt.
Knowledge attached to the job adds saved document content. Include Files adds matching input files, with optional name and file-type filters. Supported file formats depend on the selected model. Images from attached knowledge can also supply visual context.
Declared result
An output schema names the fields and types the model should return. A summary can be a string. An invoice total can be a number. The fields become action.* values, which the job maps into its output event.
Without an output schema, the generated text appears in execution history but does not create named action.* values. To pass generated text to another job, declare a string field such as summary and map {{ action.summary }} into the output event.
Extract an invoice total
For example, include an invoice PDF and request its total and currency. This configuration declares two values that a following database action can store. The example uses GPT-6 Luna.
{
"llm": "openai/gpt-6-luna",
"prompt": "Read the attached invoice. Return the final amount due as total and its three-letter currency as currency.",
"includeFiles": true,
"fileTypeFilter": [
"application/pdf"
],
"outputSchema": {
"total": {
"type": "number"
},
"currency": {
"type": "string"
}
}
}The output event can map total to {{ action.total }} and currency to {{ action.currency }}.
Model and prompt settings
Model & prompt settings, beside the prompt in the job’s Action section, opens the Model & prompt dialog. Its Model & prompt tab holds the Prompt, the attached Knowledge documents and the settings below. Save in the dialog applies the changes to the job, and saving the job stores them.
- Model
- The model that answers. The model decides which file formats, temperature and reasoning settings are available.
- Reasoning effort
- The reasoning levels the selected model supports, plus OAIZ default. A model without reasoning shows Not supported. It is separate from temperature.
- Temperature
- A value from 0 to 1, shown when the model supports it. Lower values reduce variation between responses.
- Timeout
- The job’s time limit in seconds, from 10 to 3600.
The Skills tab adds instructions from the OAIZ skill catalog. They are appended after the prompt when the job runs, and references such as {{ properties.body }} inside them are not filled in. The Integration tools tab lists tools from your integrations. Adding one changes the job to an Agent and keeps its prompt, model, skills and output schema.
LLM call itself calls no tools. Use Agent when the task must fetch more data or take action in a connected service. Tests still make real model calls when event publication or integration writes are off.