OpenAI: Make a decision
Ask OpenAI a yes/no, multiple-choice or rating question about text or images in a workflow, and branch on an answer that comes with probabilities.
OpenAI: Make a decision
Applies to: Make a decision 1.0.0
Ask OpenAI one question about a text, images or both, and get a bounded answer you can branch on. The step uses the OpenAI Decisions API with the gpt-6-luna model, so it never writes free text. Every answer comes with probabilities, so you decide how sure the model has to be.
Overview
Make a decision is an action of the OpenAI integration. Choose the kind of answer in Decision:
- Yes or no: how likely it is that the question is true for the content, turned into yes or no with a threshold.
- Pick one option: which of your options fits the content best.
- Score on a scale: where the content falls on your levels, from lowest to highest.
Settings
| Field | Description | Default |
|---|---|---|
| Account | The OpenAI connection whose API key the request uses. | |
| Decision | Yes or no, Pick one option or Score on a scale. | Yes or no |
| Question | What to decide about the content, for example Is the customer asking for a refund? Add any guidance the model should follow, such as what each option means. | |
| Text (optional) | The text to decide on. Required unless you add images. | |
| Images (optional) | Images to decide on, together with the text. Add a public image URL (http://, https:// or data:), or map a file field from a previous step, one row per source. A mapped field that holds several files sends all of them. OpenAI downloads the images, so they must be reachable from the internet. A row that is empty after variables are filled in is skipped. | |
| Options | Shown for Pick one option. The answers to pick from, 2 to 255, each unique. | |
| Levels | Shown for Score on a scale. The levels from lowest to highest, for example Low, Medium, High. 2 to 10, each unique. | |
| Threshold (optional) | Shown for Yes or no. The answer is yes when the probability is at least this value, between 0 and 1. | 0.5 |
Output
The output depends on the Decision:
| Decision | Output |
|---|---|
| Yes or no | result is true when probability reaches the Threshold. probability is how likely the question is true, from 0 to 1. |
| Pick one option | choice is the best option, exactly as you wrote it. confidence is how sure the model is, from 0 to 1. probabilities lists each option as value and probability. |
| Score on a scale | score is the probability-weighted position on the scale, where the lowest level is 0. It can fall between two levels, for example 1.35. label is the single most likely level (the lower one on a tie). confidence is how sure the model is, from 0 to 1. probabilities lists each level as level (its position from 0), label and probability. |
To act on a yes or no decision, add a Branch step on result. To act only when the model is very sure, raise the Threshold, or compare probability or confidence with your own value.
When the step fails
The step fails, with the message in the run details, when:
- a setting is missing or out of range:
- “An OpenAI connection is required.”
- “A question is required.”
- “Add the text or at least one image to decide on.”
- “Image URLs must start with http://, https:// or data:, but got “ftp://example.com/photo.png”.” The message names the first image that was rejected.
- “The threshold must be between 0 and 1.”
- “Add between 2 and 255 options.” or “Add between 2 and 10 levels.”
- “The options must not be empty.” or “The levels must not be empty.”
- “The options must be unique.” or “The levels must be unique.”
- OpenAI declines to answer, for example because of its content policy: “OpenAI declined to answer the question. Rephrase the question or the content and try again.”
- OpenAI returns no answer: “OpenAI returned no answer to the question.”
- OpenAI rejects the request, for example for a wrong API key. The message shows OpenAI’s error.