Yao Decision

Decision Tools

Use these tools when you need structured, typed decisions rather than free-form LLM text. Decision models return machine-readable answers with probabilities and confidence scores.

decision_decide

Send a state (context) and one or more typed questions to a decision model.

Question Types

Basic usage (choice + noul):

bashCopy code
tai tool decision_decide --state "Customer says: my billing is wrong, I was charged twice" --questions '{"department":{"type":"choice","instructions":"Route to the right department","criteria":{"billing":"Payment issues","technical":"Product bugs","sales":"Purchase inquiries"}},"is_urgent":{"type":"noul","instructions":"Is this urgent?"}}'

With score question:

bashCopy code
tai tool decision_decide --state "The agent resolved the issue in 2 minutes with a clear explanation" --questions '{"quality":{"type":"score","instructions":"Rate the support quality","criteria":["Poor","Below average","Average","Good","Excellent"]}}'

With explicit provider:

bashCopy code
tai tool decision_decide --state "some context" --questions '{"q1":{"type":"noul","instructions":"Is this positive?"}}' --provider t862699292876.typesafe

With model override:

bashCopy code
tai tool decision_decide --state "some context" --questions '{"q1":{"type":"noul","instructions":"Is this positive?"}}' --model jev-latest

Batch decisions (multiple states, same questions):

bashCopy code
tai tool decision_decide --states '["Customer A says billing is wrong","Customer B loves the product","Customer C wants a refund"]' --questions '{"sentiment":{"type":"choice","instructions":"Classify sentiment","criteria":{"positive":"Happy","negative":"Unhappy","neutral":"Neither"}}}'

Reading from files:

bashCopy code
tai tool decision_decide --state_file customer_message.txt --questions_file routing_questions.json

Response Format

Single state:

jsonCopy code
{
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": {"billing": 0.92, "technical": 0.05, "sales": 0.03},
      "confidence": 0.89
    },
    "is_urgent": {
      "type": "noul",
      "noul": 0.73
    }
  },
  "model": "jev-1.13.0",
  "usage": {"input_tokens": 245, "output_tokens": 38}
}

Note: model in the response is the resolved concrete model ID (e.g. jev-1.13.0), not the alias you requested (e.g. jev-latest).

Batch states:

jsonCopy code
{
  "results": [
    {"state_index": 0, "answers": {...}, "model": "jev-1.13.0", "usage": {...}},
    {"state_index": 1, "answers": {...}, "model": "jev-1.13.0", "usage": {...}},
    {"state_index": 2, "error": "timed out after 60s"}
  ]
}

decision_providers

List available decision providers and their models.

bashCopy code
tai tool decision_providers

Returns a list of providers with available models, connector IDs, and availability status.

Constraints

• state and states are mutually exclusive; provide one or the other.
• Each question must have type (choice/score/noul) and instructions.
• criteria is required for choice (map) and score (array); noul must not have criteria.
• Batch states must not contain empty strings or null values.
• timeout defaults to 60s; values ≤ 0 are treated as the default.
• Batch mode runs up to 5 concurrent calls; timeout applies to the entire batch.
• Response model is the resolved concrete model ID (e.g. jev-1.13.0), not the requested alias.