AI+法律科技落地：智能合同审查与法律文书生成

from unstructured.partition.pdf import partition_pdf
from paddleocr import PaddleOCR
from langchain.text_splitter import RecursiveCharacterTextSplitter

class ContractPreprocessor:
    def __init__(self):
        self.ocr = PaddleOCR(use_angle_cls=True, lang='ch')
        self.splitter = RecursiveCharacterTextSplitter(
            chunk_size=512, chunk_overlap=50,
            separators=["\n第.*条", "\n\n", "\n", "。"]
        )

    def process(self, file_path: str) -> list[dict]:
        elements = partition_pdf(file_path, strategy="hi_res")
        clauses = []
        current_clause = {"number": "", "title": "", "content": ""}

        for el in elements:
            text = str(el).strip()
            if re.match(r'^第[一二三四五六七八九十\d]+条', text):
                if current_clause["content"]:
                    clauses.append(current_clause.copy())
                parts = re.match(r'^(第.+?条)\s*(.*)', text)
                current_clause = {"number": parts.group(1), "title": parts.group(2), "content": ""}
            else:
                current_clause["content"] += text + "\n"

        if current_clause["content"]:
            clauses.append(current_clause)
        return clauses

from pydantic import BaseModel
from enum import Enum

class RiskLevel(str, Enum):
    HIGH = "高风险"
    MEDIUM = "中风险"
    LOW = "低风险"
    NONE = "无风险"

class ClauseAnalysis(BaseModel):
    clause_number: str
    clause_type: str          # 违约责任/知识产权/保密/管辖等
    risk_level: RiskLevel
    risk_description: str
    suggestion: str
    related_laws: list[str]   # 相关法条引用

RISK_PROMPT = """你是一名资深合同审查律师。请分析以下合同条款，评估风险等级。

条款编号：{clause_number}
条款内容：{clause_text}

审查标准：
1. 是否存在霸王条款
2. 违约责任是否对等
3. 是否违反强制性法规
4. 关键权利义务是否模糊
5. 争议解决条款是否合理

请以JSON格式返回，包含：clause_type, risk_level, risk_description, suggestion, related_laws"""

async def analyze_clause(client, clause: dict) -> ClauseAnalysis:
    response = await client.beta.chat.completions.parse(
        model="gpt-4o",
        messages=[{"role": "user", "content": RISK_PROMPT.format(**clause)}],
        response_format=ClauseAnalysis
    )
    return response.choices[0].message.parsed

from langchain_community.vectorstores import Chroma
from langchain_openai import OpenAIEmbeddings
from langchain.chains import RetrievalQA

class LegalDocumentGenerator:
    def __init__(self, db_path: str = "./legal_vectordb"):
        self.embeddings = OpenAIEmbeddings(model="text-embedding-3-large")
        self.vectordb = Chroma(persist_directory=db_path, embedding_function=self.embeddings)
        self.retriever = self.vectordb.as_retriever(search_kwargs={"k": 5})

    async def generate_brief(self, case_facts: str, doc_type: str = "起诉状") -> str:
        # 检索相关法条和判例
        relevant_docs = self.retriever.invoke(case_facts)
        context = "\n\n".join([d.page_content for d in relevant_docs])

        prompt = f"""基于以下法律依据，生成{doc_type}：

案件事实：{case_facts}

参考法条和判例：
{context}

要求：
1. 严格按照《民事诉讼文书样式》格式
2. 引用法条需标注具体条款号
3. 事实陈述需有证据对应
4. 诉讼请求需明确具体金额"""

        response = await self.client.chat.completions.create(
            model="gpt-4o", messages=[{"role": "user", "content": prompt}]
        )
        return response.choices[0].message.content

import spacy

# 基于法律语料微调的NER模型
nlp = spacy.load("zh_legal_ner_v2")

def extract_legal_entities(text: str) -> dict:
    doc = nlp(text)
    entities = {"当事人": [], "金额": [], "日期": [], "法条": [], "法院": []}
    label_map = {"PARTY": "当事人", "MONEY": "金额", "DATE": "日期",
                 "LAW_REF": "法条", "COURT": "法院"}
    for ent in doc.ents:
        if ent.label_ in label_map:
            entities[label_map[ent.label_]].append(ent.text)
    return entities

class LegalFactChecker:
    def verify_citation(self, citation: str) -> bool:
        """验证法条引用是否存在于数据库中"""
        result = self.law_db.query(
            "SELECT * FROM laws WHERE full_text LIKE ?", f"%{citation}%"
        )
        return len(result) > 0

    def verify_case(self, case_number: str) -> dict | None:
        """验证案例编号是否真实存在"""
        return self.case_db.get_case(case_number)