Prompt工程四轮迭代实录：从45%到92%的抽取准确率跃迁

from openai import OpenAI
client = OpenAI(api_key="sk-xxxx")

def extract_v1(text: str) -> str:
    prompt = f"""从以下判决书中提取：原告、被告、案由、判决金额、代理律师。
    判决书内容：{text}"""
    resp = client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.0,
        max_tokens=800
    )
    return resp.choices[0].message.content

from pydantic import BaseModel, Field
import json

class Judgment(BaseModel):
    plaintiff: str = Field(description="原告全称")
    defendant: str = Field(description="被告全称")
    cause: str = Field(description="案由，如民间借贷纠纷")
    amount: float = Field(description="判决支付金额（只含赔偿/货款，不含诉讼费）")
    lawyer: str | None = Field(description="代理律师姓名，没有则null")

def extract_v2(text: str) -> Judgment:
    schema_json = Judgment.model_json_schema()
    prompt = f"""你是法律信息抽取助手。根据判决书生成JSON，严格遵守schema：{schema_json}
    注意：金额只取"判决如下"之后出现的首个具体数字，排除"案件受理费""保全费"。
    判决书：{text[:1500]}"""  # 截断节省token
    resp = client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18",
        messages=[{"role": "user", "content": prompt}],
        temperature=0.0,
        response_format={"type": "json_object"},
        max_tokens=600
    )
    raw = resp.choices[0].message.content
    # 正则兜底：如果JSON解析失败，尝试提取字段
    try:
        return Judgment(**json.loads(raw))
    except:
        # 简单正则匹配金额
        import re
        m = re.search(r"([0-9,，]+\.?\d*)[元万]", text)
        return Judgment(plaintiff="", defendant="", cause="", amount=float(m.group(1).replace(",","")) if m else 0.0, lawyer=None)

few_shot_examples = [
    {
        "text": "反诉原告（本诉被告）甲公司诉称...本诉原告乙公司辩称...",
        "answer": {"plaintiff": "乙公司", "defendant": "甲公司", "cause": "买卖合同纠纷", "amount": 0.0, "lawyer": null}
    },
    {
        "text": "判决如下：一、被告丙公司于本判决生效后七日内赔偿原告丁某医疗费8,322.5元；二、驳回其他诉讼请求。",
        "answer": {"plaintiff": "丁某", "defendant": "丙公司", "cause": "生命权健康权纠纷", "amount": 8322.5, "lawyer": null}
    }
]

def build_prompt_v3(text: str) -> str:
    ex_text = ""
    for e in few_shot_examples:
        ex_text += f"输入：{e['text']}\n输出：{json.dumps(e['answer'], ensure_ascii=False)}\n\n"
    return f"""你是法律文书抽取引擎。规则：
    1. 原告=本诉原告（即使原文写"反诉原告（本诉被告）"）。
    2. 代理律师仅指原告委托的诉讼代理人，被告律师填null。
    3. 金额取判决主文第一条的赔偿数额，不含利息、诉讼费。
    示例：{ex_text}
    请抽取：{text[:1200]}"""

def extract_v4(text: str) -> Judgment:
    # 步骤1：粗抽取
    pre_text = text.split("判决如下")[0][:800]
    prompt1 = f"抽取JSON：{pre_text}"
    j1 = json.loads(call_gpt(prompt1))

    # 步骤2：校验律师字段
    check_prompt = f"""抽取结果：{json.dumps(j1, ensure_ascii=False)}
    原文片段：{pre_text}
    如果"lawyer"对应的姓名在原文中不是紧邻"原告委托代理人"或"原告代理人"，则把lawyer改为null。
    只回答修正后的JSON。"""
    j2 = json.loads(call_gpt(check_prompt))

    # 步骤3：金额二次确认 - 从判决主文里找
    if "判决如下" in text:
        main_text = text.split("判决如下")[1][:200]
        confirm_prompt = f"提取第一个赔偿金额数字（不含费）：{main_text}"
        amount = float(call_gpt(confirm_prompt))
        j2["amount"] = amount
    return Judgment(**j2)