Decision

The Decision node asks a decision model typed questions about the run's current state and returns probabilities. It does not generate text. Reach for it when a workflow needs a judgment it can branch on — routing, triage, ranking, verification, or gating a step that cannot be undone.

Decision models (also called System One models) are a different kind of model from an LLM. There is no prompt, no system instruction, no temperature and no tool calling. You send a state plus a map of questions, and each answer comes back as a number with a probability distribution behind it. Because an answer is only a few tokens, these models are fast and cheap compared with asking an LLM to produce JSON and parsing it.

For text generation, use the LLM node. For a model that calls tools, use the Agent Node.

Overview

Parameters

State

State is the situation the model judges. It accepts a string, an object or an array.

A field holding exactly one expression keeps its own type, so $input sends the object itself rather than its text. A template with an expression inside it (Ticket: $input.text) resolves to a string. Give each question enough to answer with: source text, identities, relationships, policies and current facts.

Question types

Each question has an id, a type, instructions, and criteria that depend on the type.

The id is how your workflow reads the answer back. It is never sent to the model, so the whole meaning has to live in instructions. Writing id: is_urgent with instructions: "Is it?" gives the model nothing to work with.

Noul

Whether a condition holds. Returns the probability that the answer is yes, from 0 to 1.

Criteria are optional: Yes means and No means describe what each side covers. A value near 0.5 means the model finds yes and no about equally likely — not "medium intensity".

{
  "id": "is_urgent",
  "type": "noul",
  "instructions": "Does this message convey urgency?",
  "criteriaTrue": "Explicitly time-sensitive or blocking work",
  "criteriaFalse": "No urgency expressed"
}

Read it with $triage.answers.is_urgent.noul.

Choice

Picks one option from a set you define. Returns the chosen option, the full distribution, and a confidence derived from it.

Add an option covering "none of these" whenever nothing may fit — the model cannot choose an option you did not offer.

{
  "id": "department",
  "type": "choice",
  "instructions": "Which team should handle this request?",
  "options": [
    { "key": "billing", "description": "Payments, invoicing, refunds" },
    { "key": "technical", "description": "Bugs, outages, integrations" },
    { "key": "other", "description": "Nothing above fits" }
  ]
}

Read it with $triage.answers.department.choice, .probabilities.billing and .confidence.

Score

Rates the state along an ordered rubric. Returns a probability-weighted value that can land between levels, plus the legend and the distribution.

Levels run lowest first, at least two and at most ten. Each level should describe a concrete situation and stand on its own.

{
  "id": "frustration",
  "type": "score",
  "instructions": "How frustrated is the customer?",
  "levels": ["Calm", "Frustrated", "Very angry"]
}

Read it with $triage.answers.frustration.score, .legend and .confidence.

Generate with AI

Generate with AI turns a plain sentence into a suggested state template and a set of question rows. Describe what you want judged — "triage an incoming support ticket: how urgent it is, which team owns it, and how frustrated the customer sounds" — and review what comes back before adding it to the node.

The drafting is done by an LLM, so the dialog asks for an LLM credential. The decision model answers the questions later, when the workflow runs. Generated ids are de-duplicated against the rows the node already holds, so generating twice never overwrites an existing answer key.

Each generation is recorded in the Traces tab under the Decision Questions source, alongside the decision.systemone traces written by the node itself at run time.

Custom request body

Turn on Custom request body to send JSON you write yourself instead of the form. Use it for an endpoint whose contract differs from the one the form builds.

The editor is seeded from the current form, and expressions inside it are resolved before the request is sent. While the toggle is on, the State field and the question rows are ignored. The response is passed through unchanged either way, so a different contract simply produces a different shape for your expressions to read.

Output

The node returns exactly what the provider returned. Nothing is renamed or flattened.

{
  "model": "jev-1.13.0",
  "answers": {
    "is_urgent": { "type": "noul", "noul": 0.95 },
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": { "billing": 0.88, "technical": 0.12, "other": 0.0 },
      "confidence": 0.81
    },
    "frustration": {
      "type": "score",
      "score": 1.05,
      "legend": { "0": "Calm", "1": "Frustrated", "2": "Very angry" },
      "probabilities": { "0": 0.0, "1": 0.95, "2": 0.05 },
      "confidence": 0.92
    }
  },
  "usage": { "input_tokens": 296, "output_tokens": 20 }
}

Branching on an answer

A Switch node routes on the chosen option, and a Condition node can send low-confidence cases somewhere a person will see them:

Input → Decision (triage) → Condition ($triage.answers.department.confidence > 0.7)
                              ├─ true  → Switch on $triage.answers.department.choice
                              └─ false → Send Email (human review)

A typed answer guarantees the shape, not the truth. Pick thresholds against your own data and consequences rather than copying a number from an example.

Credential

Create a Decision Model credential in the Credentials tab with:

• Base URL — the endpoint, for example https://api.typesafe.ai. Heym appends /v1/systemone.
• API Key — optional. Endpoints that need no key can be left blank.

Editing a credential keeps the stored key when the key field is left blank; retype it only to replace it.

Heym blocks outbound requests to private and loopback addresses by default, so a decision endpoint you run yourself on localhost or a private range needs HEYM_HTTP_ALLOW_PRIVATE_URLS=true.

Decision models do not appear in the LLM pricing tables Heym syncs, so the Cost column in Traces stays empty for these calls. Token counts are still recorded, and you can add your own price with a pricing override.

Errors

Related

• LLM Node – Generate text instead of a judgment
• Agent Node – Gather evidence with tools, then judge it here
• Switch Node – Route on the chosen option
• Condition Node – Gate on confidence
• Node Types – Overview of all node types
• Credentials – Credential types and fields
• Traces Tab – Inspect decision calls and token usage