# Configure a decision model in Discourse AI

**URL:** <https://meta.discourse.org/t/configure-a-decision-model-in-discourse-ai/416486>\
**Category:** Site Management\
**Tags:** how-to, ai\
**Created:** [4. November 2026 um 01:05 UTC](https://meta.discourse.org/t/configure-a-decision-model-in-discourse-ai/416486 "2026-11-04T01:05:32Z")\
**Posts on this page:** 1\
**Page:** 1

## AI-generated summary

This guide details how to configure and test a decision model in Discourse AI using the hosted D3 service. Decision models are distinct from language models; they classify input (yes/no, choice, or score) rather than generating conversational text. Configuration requires Administrator access, the `discourse_ai_enabled` flag, and a valid credential for the decision endpoint. Note that a Workflows plugin or separate language model is not required solely for configuration and testing.

To begin, administrators should create an API key via the [D3 Decision API](https://decision.vllm-sr.ai/). The D3 service is free but may use submitted data for training and evaluation, so it is strictly recommended to use invented, non-confidential examples for testing. Once the key is obtained, it must be saved in **Admin → Plugins → AI → Credentials** under a descriptive name.

The guide explains that decision models are used in specific features like sentiment analysis, emotion detection, and combined-search intent. These features have their own prerequisites and settings (e.g., `ai_sentiment_enabled`, `ai_sentiment_emotion_enabled`, and `ai_ask_ai_search_intent_decision_model`). Configuring the model does not automatically enable these features; administrators must explicitly activate them. The D3 example should only be used with synthetic data unless the provider's policies are reviewed and approved for real forum content.

A troubleshooting section addresses common issues:
- **Missing Models**: Ensure Discourse AI is enabled and check the unified Models interface.
- **Authentication Failures**: Verify the correct credential is selected.
- **HTTP/Response Errors**: Check the endpoint URL, protocol, and model identifier.
- **Image Failures**: Ensure the endpoint supports image input or disable "Supports image input" for text-only models.
- **No Action on Forum**: Remember that testing does not activate features; you must save the connection and configure the consuming feature separately.

For debugging, admins can review request details in **Admin → Plugins → AI → Logs**, but must treat logs as sensitive and avoid pasting credentials or private data in support forums. Additional resources and links to related documentation are provided at the end of the post.

<div class="post-metadata">

**Author:** ![meta-bot](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/meta-bot/32/582912_2.png) [@meta-bot](https://meta.discourse.org/u/meta-bot)\
**Post date:** [4. November 2026 um 01:05 UTC](https://meta.discourse.org/t/configure-a-decision-model-in-discourse-ai/416486/1 "2026-11-04T01:05:32Z")

</div>

> 🔖 This guide explains how to configure and test a decision model in Discourse AI, using the hosted D3 service as an example.
> 
> 🙋 Required user level: Administrator

A decision model classifies an input instead of writing a conversational answer. You can use its structured results in supported Discourse AI features and workflows.

Discourse supports three question types:

| Type | Result |
| --- | --- |
| Yes or no | A probability between 0 and 1 |
| Choice | A selected option, probabilities and confidence |
| Score | A score across ordered levels, with probabilities and confidence |

A decision model is separate from the language models used for AI bot replies, summaries and Helper. Configuring one does not replace your language model or automatically enable a feature.

## Before you start

You need:

- Discourse AI enabled through `discourse_ai_enabled`.
- An administrator account.
- A compatible decision-model endpoint and any credential it requires.

You don’t need the Workflows plugin or a separate language model to configure and test a decision model. A workflow requires the Workflows plugin; other features have their own prerequisites.

## 1. Obtain a credential for the hosted D3 example

The [D3 Decision API](https://decision.vllm-sr.ai/) offers free hosted access with an API key. Review its [terms](https://decision.vllm-sr.ai/terms) and [privacy notice](https://decision.vllm-sr.ai/privacy) before creating a key. Creating a key requires accepting the terms, including the age requirement. No account is required.

⚠ The hosted D3 service may store submitted text, images and responses, use them for model training and evaluation, and publish de-identified data. No request header disables this. Use invented examples for testing. Don’t connect private, confidential or regulated forum content to this service.

1. Open the D3 service and use its API-key creation control.
2. Copy the key when it is displayed. The service only shows it once.
3. In Discourse, open **Admin → Plugins → AI → Credentials**.
4. Click “New credential”.
5. Enter a descriptive name and paste the key into “Value”.
6. Click “Save”.

See [Managing AI credentials](https://meta.discourse.org/t/397407) for credential rotation and deletion. Keep the key out of posts, screenshots and client-side code.

Free access is subject to the provider’s quotas, terms and availability. Check the [provider documentation](https://decision.vllm-sr.ai/docs) for current limits rather than treating it as an unlimited service.

## 2. Add the decision model

1. Open **Admin → Plugins → AI → Models** , or visit `/admin/plugins/discourse-ai/ai-models`.
2. Click “New model”.
3. Choose “Decision model”.
4. Find the “d3” template and click “Set up”.

 ![Decision model templates with the d3 option](https://global.discourse-cdn.com/meta/original/4X/9/e/2/9e242d6405c0414dab4a4618727d7cf7a9c17dc0.png)

The template is a starting point, not a provisioned service. It leaves the endpoint URL empty. Its description refers to self-hosting, but you can supply a compatible hosted endpoint.

1. Fill in the following fields:

| Field | Hosted D3 example |
| --- | --- |
| Name | `D3 documentation test`, or another descriptive display name |
| Provider | “System One” |
| Model | `vllm-sr/d3-ultra` |
| Endpoint URL | `https://decision.vllm-sr.ai/v1/systemone` |
| Credential | The credential you created in the previous section |
| Custom headers | Leave empty for this example |
| Supports image input | Enable for this model if you want to test image support |

Use the complete inference URL. Discourse sends requests to the URL you enter; it doesn’t append `/v1/systemone` for you.

The preset’s `vllm-sr/d3` model name is an alias at this hosted service. Using the explicit `vllm-sr/d3-ultra` identifier makes the example’s model choice clear. Check the provider’s model list if an identifier becomes unavailable.

 ![D3 configuration with a named credential and image input enabled](https://global.discourse-cdn.com/meta/original/4X/e/5/9/e59c5ea5c332fe8e587580a0d0d27aa9f6c3e71d.png)

For other providers, choose their matching template and supply their endpoint, model identifier and credential. A System One endpoint is not an OpenAI Chat Completions endpoint.

## 3. Test and save the connection

1. Click “Run test”. This tests the current form values, including unsaved changes.
2. Check the result below the form.

The built-in test sends the invented greeting “Hello! How are you doing today?” and asks whether it is a greeting. If “Supports image input” is enabled, it also sends a generated red square and checks the image response.

A successful result looks like this:

 ![Successful D3 connection test](https://global.discourse-cdn.com/meta/original/4X/d/5/4/d54e5d4744c2f524130387b87248179c551075c6.png)

The displayed percentage is the greeting’s yes probability, rounded to a whole percentage. Success confirms a valid response and, when enabled, a basic image check. The test does not require a minimum greeting probability, so success is not an accuracy rating or a guarantee about your own classification tasks.

1. Click “Submit” to save the model. “Run test” alone does not save it.
2. Confirm that Discourse shows “Decision model saved”. The model is available from the Models list for later editing.

Testing makes a real request to the configured provider and creates local AI audit records. Other providers may charge for requests.

### Example classification result

In a synthetic test through Discourse’s decision-model client, the hosted D3 model received this invented message:

> Hello! Thank you for helping me today.

The request asked three questions about the same message. These are the actual returned results, rounded for readability:

| Question | Choices or levels | Result |
| --- | --- | --- |
| Is the message a greeting? | Yes or no | Yes probability: 99.30% |
| What is the main purpose of the message? | `greeting`, `billing`, `technical support` | `greeting`, with probability 99.999% |
| How polite is the message? | `rude`, `neutral`, `polite` | `polite`, with probability 99.924% |

These are classification results, not a generated reply. A consuming feature can use such a result to choose what happens next. This test called the decision-model client directly; it did not activate a workflow, route a post or enable sentiment analysis. The admin “Run test” button runs the fixed greeting and optional image checks described above, not these three custom questions. Results on one invented message do not establish accuracy on your forum’s content.

## 4. Assign the model to a supported feature

Saving the connection does not send your forum content for classification. You must separately select the model in a consuming feature.

For this hosted D3 example, leave consumers unconfigured. Enabling one can send real forum content or search queries to the provider. Permission to run synthetic tests does not cover that data use.

| Use case | Where the model is selected |
| --- | --- |
| Workflow classification | The “Classify” node in Discourse Workflows |
| Sentiment classification | Set `ai_sentiment_sentiment_classification_strategy` to `decision_model`, then select the model in `ai_sentiment_sentiment_decision_model` |
| Emotion classification | Set `ai_sentiment_emotion_classification_strategy` to `decision_model`, then select the model in `ai_sentiment_emotion_decision_model` |
| Experimental combined-search intent | Select it in `ai_ask_ai_search_intent_decision_model` when the combined-search upcoming change is available and enabled |

Each feature also needs its own enablement and configuration. Sentiment uses `ai_sentiment_enabled`; emotion uses `ai_sentiment_emotion_enabled`. Their model selectors appear after the feature is enabled and its classification strategy is set to `decision_model`. A search-intent decision model helps select the search experience; it does not generate the AI answer.

For the hosted D3 example, stop after the synthetic connection test unless your intended data use complies with the provider’s policy. Use an appropriately reviewed provider or self-hosted endpoint for content you cannot share under those terms.

## Troubleshooting

| Problem | What to check |
| --- | --- |
| You can’t find decision models | Check that Discourse AI is enabled and your release includes decision-model support. In the unified interface, use Models, not a separate Decision models tab. |
| Authentication fails | Select the correct credential. The D3 hosted service requires a key even though access is free. |
| The endpoint returns an HTTP error or invalid response | Check the complete endpoint URL, provider protocol and model identifier. Review the provider status and error message. |
| A provider limit is reached | Follow its retry guidance and quotas. Don’t create additional keys to bypass limits. |
| The image test fails | Check that the selected model and endpoint support images. Disable “Supports image input” for a text-only endpoint and run the test again. |
| The test passes, but nothing happens on the forum | Save the connection, then configure a supported consuming feature. Testing does not activate it. |

For request details, open **Admin → Plugins → AI → Logs**. Treat logs as sensitive: they can include submitted text and responses. Never paste credentials or private request contents into a support topic.

## Additional resources

- [TypeSafe Jev and other System One decision models](https://meta.discourse.org/t/414738)
- [Managing AI credentials](https://meta.discourse.org/t/397407)
- [Discourse AI - Sentiment](https://meta.discourse.org/t/259599)
- [Experimenting with combined AI search and decision models](https://meta.discourse.org/t/415171)
- [API Docs · Decision 3.0](https://decision.vllm-sr.ai/docs)
- [Terms · Decision API](https://decision.vllm-sr.ai/terms)
- [Privacy · Decision API](https://decision.vllm-sr.ai/privacy)