> For the complete documentation index, see [llms.txt](https://docs.parasail.io/parasail-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.parasail.io/parasail-docs/guides/decision-models.md).

# Decision Models

Configure decision models and test System One requests in the Parasail playground.

Decision models answer specific questions about a shared context rather than generate a conversation. They can classify a support ticket, rate its urgency, or estimate whether it needs a follow-up. The [System One API](https://docs.parasail.io/parasail-docs/api-reference/systemone.md) returns typed answers and probabilities in one JSON response.

## Choose the right model and endpoint

Use a model that serves `POST /v1/systemone`. A model being available through Parasail doesn't mean it supports every API. Check its documented capabilities and the model mode shown in the portal; don't infer support from its name.

Use the served model ID or alias from the deployment's connection example. It can differ from the Hugging Face repository containing the weights. See the [models endpoint](https://docs.parasail.io/parasail-docs/api-reference/models-endpoint.md) for model discovery, but verify System One support separately.

For a generative model that should return JSON, use [structured output](https://docs.parasail.io/parasail-docs/guides/structured-output.md) with chat completions instead. Setting a chat response format doesn't turn a generative model into a System One decision model.

## Configure a dedicated deployment

1. Open your [Dedicated Instances](https://docs.parasail.io/parasail-docs/products/overview-1.md) deployment configuration.
2. Use a model server that implements `/v1/systemone`. If you use a custom container, its inference server must expose that route.
3. Expand **Advanced configuration**, then **Model mode**.
4. Select **Decision (System One)** and save or deploy the configuration. For programmatic deployment configuration, the corresponding `engineTask` value is `SYSTEMONE`.
5. Wait for the deployment to become available before sending requests.

Select this mode explicitly: automatic detection doesn't probe System One. The mode selects the appropriate health check and usage examples; it doesn't add an endpoint to a server that lacks it.

If **Decision (System One)** isn't available in your portal, contact support to confirm deployment support. Use the [HTTP examples](https://docs.parasail.io/parasail-docs/api-reference/systemone.md#send-a-request) for an existing System One deployment while the portal controls are unavailable.

## Try a decision in the playground

The following steps apply to the portal's decision-model playground:

1. Open **Models**, choose an available decision model, then select **Try** or open its details and select **Playground**. For a dedicated deployment, open its **Playground** tab.
2. Select **Try it out**. Decision models show a JSON request editor rather than a chat input.
3. Enter a `state` and a named `questions` object. For example, replace the editor contents with:

```json
{
  "state": "Our checkout has returned 500 errors. Customers cannot complete purchases.",
  "questions": {
    "category": {
      "type": "choice",
      "instructions": "Which team should review this ticket?",
      "criteria": {
        "engineering": "Software errors or broken features",
        "billing": "Payments, invoices, or refunds",
        "account": "Login or account access"
      }
    }
  }
}
```

4. Select **Run decision**. The playground sends a non-streaming request to `/v1/systemone` and displays the full **Response JSON**.
5. Read the result under `answers.category`. Inspect the probabilities and confidence before using it in your app.

The playground adds the selected served model ID automatically and overrides any `model` value in the editor. You can include optional `images` and `keep_alive` fields if the model server supports them. Chat sampling controls aren't part of this request.

**Cancel** stops waiting for the request. **Reset** clears the result and restores the example input. Changing models or leaving **Try it out** cancels the pending browser request. Canceling a client request doesn't guarantee the server stopped processing it or that no charges apply.

You need an available endpoint and permission to use inference. If the playground asks you to add a payment method, complete billing setup first. For an offline deployment, start it and wait for it to become available. Request failures appear beside the editor; correct invalid JSON before trying again.

## Use the code examples

The **Python**, **JavaScript**, and **cURL** tabs provide System One requests for the selected model. Python uses HTTPX and JavaScript uses fetch; the OpenAI chat-completions method doesn't support this API. For complete examples, see [Send a request](https://docs.parasail.io/parasail-docs/api-reference/systemone.md#send-a-request).

Store your API key in an environment variable, keep it out of browser code, and preserve the served model ID when adapting an example. Don't replace it with a weights repository ID unless the deployment serves that exact ID.

## Interpret a decision

* **Choice:** read the selected option and the probability of each allowed option.
* **Score:** read the position on your ordered levels; a score can fall between two levels.
* **Noul:** read the probability that your statement is true, rather than a literal yes or no.

Evaluate questions and decision thresholds against examples from your workload. A high confidence value isn't proof that an answer is correct. Treat missing answers or missing usage as unknown, not as a negative decision or zero cost. Route uncertain results for review before taking consequential actions.

## Next steps

* [System One API reference](https://docs.parasail.io/parasail-docs/api-reference/systemone.md)—request fields, question types, responses, and errors
* [Dedicated Instances](https://docs.parasail.io/parasail-docs/products/overview-1.md)—deployment and lifecycle controls
* [Pricing](https://docs.parasail.io/parasail-docs/billing/pricing.md)—billing for your deployment type
