Classify, gate and score

Use this when a step needs judgement that a rule cannot make (“is this a bug or a question?”, “is this reply safe to send?”) and the answer should come back in a shape you can route on. For open-ended writing use LLM_CALL. Prerequisites: the CLASSIFY plugin is present (GET /api/plugins/CLASSIFY answers 200). Pick a backend:

SYSTEMONE (default): a System One decision model on your own Ollama (baseUrl defaults to http://localhost:11434). Answers carry calibrated probabilities and cost no provider tokens.
LLM: any chat model, through the same access as LLM_CALL. Probabilities are the model’s own estimates.

Add -H "X-API-Key: $YUKTA_API_KEY" if your server has keys.

Steps

Classify into labels and route on them

CLASSIFY picks one label from your list. Describe every label: a small model depends on those descriptions. With routeByLabel, each label becomes an output port, plus uncertain for answers below minProbability (default 0.6). Save as session.json:

session.json

{
  "sessionId": "decide",
  "name": "Decide",
  "workflows": {
    "triage": {
      "workflowId": "triage",
      "description": "Sort a message by kind",
      "nodes": [
        { "nodeId": "in", "type": "API_TRIGGER", "config": {} },
        {
          "nodeId": "kind",
          "type": "CLASSIFY",
          "config": {
            "model": "tev1:0.8b",
            "state": "payload.text",
            "labels": {
              "bug": "The user reports something that is broken or behaves wrongly",
              "feature": "The user asks for new behaviour",
              "question": "The user asks how something works"
            },
            "routeByLabel": true,
            "minProbability": 0.6
          }
        },
        { "nodeId": "bugs", "type": "CONSOLE_TERMINAL", "config": {} },
        { "nodeId": "features", "type": "CONSOLE_TERMINAL", "config": {} },
        { "nodeId": "questions", "type": "CONSOLE_TERMINAL", "config": {} },
        { "nodeId": "unsure", "type": "CONSOLE_TERMINAL", "config": {} }
      ],
      "edges": [
        { "source": "in", "target": "kind" },
        { "source": "kind", "target": "bugs", "sourcePort": "bug" },
        { "source": "kind", "target": "features", "sourcePort": "feature" },
        { "source": "kind", "target": "questions", "sourcePort": "question" },
        { "source": "kind", "target": "unsure", "sourcePort": "uncertain" }
      ]
    }
  }
}

The model name tev1:0.8b is an example; use the decision model you pulled. state is what the model sees (default the whole payload, cut to 15,000 characters). To use a chat model instead, add "backend": "LLM", "provider": "anthropic", "apiKey": "decrypted:anthropic" and its model.

curl -X POST localhost:9858/api/sessions -H 'Content-Type: application/json' -d @session.json
curl -X POST localhost:9858/api/workflow/start -H 'Content-Type: application/json' \
  -d '{"sessionId":"decide","workflowId":"triage","payload":{"text":"The save button crashes the app"}}'

Gate on a yes or no

GATE asks a yes/no question and leaves on pass when the probability of yes reaches threshold (default 0.5), else reject. Put one in front of a risky node:

{
  "nodeId": "safe-to-send",
  "type": "GATE",
  "config": {
    "backend": "LLM", "provider": "anthropic", "model": "claude-sonnet-5-5", "apiKey": "decrypted:anthropic",
    "question": "Is this reply polite and free of personal data?",
    "state": "payload.reply",
    "threshold": 0.8
  }
}

Score on a scale

SCORE rates on ordered levels (2 to 26, lowest first). With threshold it leaves on high or low.

{ "nodeId": "urgency", "type": "SCORE",
  "config": { "model": "tev1:0.8b", "question": "How urgent is this ticket?",
              "levels": ["can wait", "this week", "today", "now"], "threshold": 2 } }

Ask several things at once, or watch a stream

ASK takes up to 64 named questions (choice, yesno, score) in one call and writes metadata.answers; route with a BRANCH on metadata['answers']['sentiment']['choice'].
ANOMALY and DEDUPE compare each message with a window of recent ones (ports normal/anomalous and unique/duplicate). By default the window is per node and per run, so a workflow started for every event never builds history; set persist: session to share it across runs.

Verify it worked

curl -s -H 'Accept: application/json' localhost:9858/api/sessions/decide/executions/$EXEC/results

The result comes from the bugs terminal, and the message metadata holds the decision: classification with choice, probability, probabilities, label and uncertain. The payload is unchanged. Send {"text":"hmm"}: a vague message should land on unsure. If the run fails with a model error, the Ollama endpoint is unreachable or the model is not pulled: the node fails and the node’s engine policy applies (Handle failures).

Variations

Where the result lives. Each plugin writes to metadata[into] (classification, gate, score, answers, anomaly, dedupe); change it with into.
Cap what the model sees. maxStateChars (max 15,000), stateKeep: tail for build logs where the cause is last, and onOverflow: fail to reject oversize input instead of cutting it.
Latency. keepAlive (such as 30m) keeps an idle Ollama model loaded; concurrency (default 16) sets calls in flight.
Gate a tool call. Put a GATE in front of MCP_TOOL_CALL asking whether the arguments the model proposed are safe: Use tools and MCP servers.

Decision plugins reference for every option, port and default.
AI and deterministic work for when to use a decision model instead of a chat model.

Build workflows

Work with AI

Supervise runs

Run and operate

Extend

Classify, gate and score

Steps

Verify it worked

Variations

Steps

Verify it worked

Variations

Related