我用5版Prompt把合同要素抽取准确率从71%做到94%，token反而降了38%

import json
from openai import OpenAI

client = OpenAI(api_key="sk-xxx")

SCHEMA = {
    "type": "object",
    "properties": {
        "contract_no": {"type": ["string", "null"]},
        "party_a": {"type": ["string", "null"]},
        "party_b": {"type": ["string", "null"]},
        "sign_date": {"type": ["string", "null"],
                      "description": "ISO 8601, YYYY-MM-DD，无法确定填null"},
        "total_amount": {"type": ["number", "null"],
                         "description": "单位元，去掉千分位和货币符号"},
        "payment_terms": {"type": ["string", "null"],
                          "description": "原文摘录，不超过80字，不要改写"},
    },
    "required": ["contract_no", "party_a", "party_b",
                 "sign_date", "total_amount", "payment_terms"],
    "additionalProperties": False,
}

SYSTEM_PROMPT = """你是合同信息抽取器。只做抽取，不做推理和总结。
规则：
1. 字段值必须能在原文中找到依据，找不到填 null，禁止编造。
2. 金额统一换算为元，去掉"人民币""¥"和千分位逗号。
3. 中文数字日期（如"二〇二四年三月五日"）转为 YYYY-MM-DD。
4. payment_terms 必须是原文连续摘录，不要概括。"""

def extract_v3(text: str) -> dict:
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        temperature=0.1,
        max_tokens=1024,
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": f"合同原文：\n{text}"},
        ],
        response_format={
            "type": "json_schema",
            "json_schema": {"name": "contract", "strict": True, "schema": SCHEMA},
        },
    )
    return json.loads(resp.choices[0].message.content), resp.usage

def extract_v5(text: str) -> dict:
    first, u1 = extract_v3(text)
    check_prompt = f"""原文：
{text}

第一次抽取结果：
{json.dumps(first, ensure_ascii=False)}

逐字段核对：值是否能在原文找到依据？金额单位是否正确？日期是否为YYYY-MM-DD？
只输出修正后的JSON，字段和类型不变。"""
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        temperature=0.0,
        max_tokens=1024,
        messages=[{"role": "user", "content": check_prompt}],
        response_format={
            "type": "json_schema",
            "json_schema": {"name": "contract", "strict": True, "schema": SCHEMA},
        },
    )
    return json.loads(resp.choices[0].message.content), (u1, resp.usage)