Structured extraction with a verification pass

Recover a value from messy text while keeping normalization and validation deterministic.

CHOICE×2

One request, 2 parallel answers (three shown). The model never performs the side effect — your code reads the confidences and decides.

HOW IT WORKS

state = {
    "document": "Invoice total: INR 1,24,500.00 payable by 14/10/26.",
    "candidate_amounts": ["1,24,500.00"],
    "candidate_dates": ["14/10/26"],
}

questions = {
    "amount_candidate": Choice(
        instructions="Which item in `candidate_amounts` is the invoice total?",
        criteria={"0": "The candidate at index 0 is the invoice total.", "none": "No candidate is the invoice total."},
    ),
    "date_candidate": Choice(
        instructions="Which item in `candidate_dates` is the payment due date?",
        criteria={"0": "The candidate at index 0 is the payment due date.", "none": "No candidate is the payment due date."},
    ),
}

After Jev selects a candidate, parse it with a locale-aware library, validate the range and currency, and reject malformed values. Do not ask Jev to emit a free-form number when a parser can do the final conversion.

Source: Pre-parsed value extraction and date extraction.

요청 1 / 답 2

이 패턴을 한 줄로