Decision Presets

Decision presets evaluate text against questions you define. Use them for classification, scoring, and yes-or-no checks that need a stable @preset model value and structured results.

Examples include triaging a support request, scoring urgency, checking whether a document meets criteria, or choosing a workflow category.

Prepare a decision model

Open Models > Decisions. Nexus recommends a compatible model and runtime for the machine.

1. Review the recommended model and runtime.
2. Select Set up.
3. Let Nexus install or start the runtime when needed.
4. Wait for the model to download and load.
5. Set it as the default decision model.

Decision support is a model capability. Compatible decision models can run through MLX, llama.cpp, Ollama, or a configured online provider. They do not appear as normal chat models because their output is an evaluation, not generated prose.

Create a decision preset

1. Open Presets > New preset.
2. Select Decision preset.
3. Choose the primary decision model.
4. Add one or more questions.
5. Choose whether to evaluate the latest user message or the whole conversation.
6. Test the draft, review probabilities and timing, then save it.

Question types include:

• Choose one for a category from a defined list.
• Score for an ordered scale.
• Yes / no for a binary check.

The editor keeps changes as a draft until you save. Testing a draft records normal usage but does not create or modify a preset.

Use inputs and templates

A preset can define typed inputs that a client supplies with preset_inputs. Inputs let one saved decision reuse a controlled set of labels, levels, or reference values without letting the caller replace the decision questions.

Supported input shapes include text, choices, ordered levels, yes/no descriptions, and JSON. Give required inputs a default when most callers should not need to send them.

Nexus binds inputs once. Supplied values are treated as data, not as another template. This prevents a value from injecting new placeholders or changing object keys.

Call the preset

Decision presets use the same model value pattern as other presets:

@preset/<slug>

They can be called through OpenAI-compatible Chat Completions. Clients that need the provider-native decision contract can use /v1/systemone instead.

Example request body:

{
  "model": "@preset/message-category",
  "messages": [{"role": "user", "content": "Can you add dark mode?"}],
  "preset_inputs": {
    "categories": ["Bug report", "Feature request", "Question"]
  }
}

Choose readable output for a direct answer or JSON when the client needs answers, confidence, and probability details.

Configure fallback models

Add up to three concrete decision models as fallbacks. Nexus tries them in order when the primary model is unavailable, loading, rate-limited, failing, or over its supported limits.

Fallbacks do not bypass policy. Every attempt uses the same client access, credentials, Data Protection settings, admission controls, and usage recording. Invalid input or a policy refusal is not retried because another model would reject the same request.

A practical order is a fast local model followed by a compatible online model.

Review uncertain answers

Optional review thresholds can flag a low-confidence choice or score, or a yes/no result near the middle. They are workflow hints, not guarantees of correctness.

Model confidence is probabilistic. Do not use a decision result as the only permission check, security control, or irreversible approval.

Privacy and limits

• Local decision models keep evaluation text on the selected Nexus machine.
• Online decision providers receive the text needed for the evaluation.
• Data Protection can inspect a decision preset before it reaches the model.
• Whole-conversation mode accepts text messages; system instructions do not replace the preset's saved questions.
• Attachments and tool messages can be rejected when the decision contract cannot represent them safely.

If Describe it in the Smart Preset creator offers local description reading, Nexus uses the prepared decision model on this machine. The description stays local.