Choice: pick one
Choice is Jev’s workhorse: select one option from a set of at most 255 options you define. It fits routing, classification, and picking — any “pick one from a closed set” judgment.
Request
Section titled “Request”{ "type": "choice", "instructions": "Which team should handle this email", "criteria": { "billing": "Payment, subscription, or invoice issues", "technical": "Bugs, integrations, API problems", "sales": "Pricing, plans, account questions" }}criteriais a map: keys are the option IDs returned to you; values are the semantic descriptions shown to the model- Cap: 255 options (official docs, verified 2026-09-22)
- Descriptions accept natural language; the official advanced guide notes structured JSON descriptions are also accepted
Answer
Section titled “Answer”{ "type": "choice", "choice": "technical", "probabilities": { "technical": 0.85, "billing": 0.15, "sales": 0.0 }, "confidence": 0.78}| Field | Meaning | Use |
|---|---|---|
choice |
The selected option ID | Drive your business branches directly |
probabilities |
Probability for every option | Runner-up fallbacks, multi-way dispatch, weighted policies |
confidence |
Distribution-shape statistic, not accuracy | Below threshold → human/fallback (below) |
Typical uses
Section titled “Typical uses”- Intent routing: dispatch each request to deterministic logic, a specialist model, or a human — the official Intent routing pattern; our walkthrough is in Model routing
- Hierarchical classification: the official cookbook runs parallel beam search over deep patent, retail, biomedical, and source-code category trees — ask the top level, expand children by probability (via the llms.txt index)
- Re-ranking: the official cookbook asks one question per query–candidate pair for 40 legal queries, lifting top-1 accuracy from 5% to 18% and top-10 from 38% to 62% (official figures)
Combining with other types
Section titled “Combining with other types”Choice pairs naturally with Noul: pick the category, then confirm “does this need human review?” — both in the same request. See Capabilities on parallel questions and the full example in the API reference.