Jev AI

Hierarchical classification

Classify into a deep taxonomy without asking one question to carry an unwieldy option list.

CHOICE×1
UNSTRUCTUREDSTATEJEVCHOICE+ confidenceCODE-OWNEDROUTINGACTESCALATEREVIEW
One request, 1 parallel answer. The model never performs the side effect — your code reads the confidences and decides.
01
HOW IT WORKS
root = Choice(
    instructions="Which top-level product area does this document belong to?",
    criteria={
        "payments": "Charges, payouts, invoices, and refunds.",
        "identity": "Login, authentication, and user verification.",
        "analytics": "Reports, dashboards, and data exports.",
        "other": "No top-level area clearly fits.",
    },
)

If the chosen branch changes the next options, build a second state/question set in code. If all questions could have been asked against the original state, prefer one fan-out request and let code ignore irrelevant answers. Use an explicit other or review leaf so the taxonomy can admit uncertainty.

Source: Hierarchical classification and when one question depends on another.

요청 1 / 답 1
이 패턴을 한 줄로