AIRegulation-DocAnalysis/backend/app/services/llm/document_summarizer.py

# src/services/llm/document_summarizer.py
"""文档摘要生成服务 - LLM生成法规文档摘要"""

from typing import Dict, Optional
from dataclasses import dataclass
from loguru import logger

from app.services.llm import get_llm_client, BaseLLMClient
from app.services.rag.prompt_templates import get_prompt_template
from app.config.settings import settings


@dataclass
class DocumentSummary:
    """文档摘要结果"""
    doc_name: str
    summary: str
    applicable_scope: str
    key_clauses: list
    key_terms: list
    compliance_points: list
    model: str
    latency_ms: int
    error: Optional[str] = None

    @property
    def is_success(self) -> bool:
        return self.error is None


class DocumentSummarizer:
    """
    文档摘要生成器

    功能：
    - 生成法规文档的核心要点摘要
    - 提取适用范围
    - 突出关键条款
    - 列出合规要点

    使用示例：
        summarizer = DocumentSummarizer()
        result = summarizer.summarize("GB 7258-2017", markdown_content)
        print(result.summary)
    """

    def __init__(
        self,
        provider: str = None,
        model: str = None,
        max_tokens: int = None
    ):
        """
        初始化摘要生成器

        Args:
            provider: LLM提供商
            model: LLM模型名称
            max_tokens: 最大输出token数
        """
        self.provider = provider or settings.llm_provider
        self.model = model or settings.llm_model
        self.max_tokens = max_tokens or settings.rag_summary_max_tokens

        # LLM客户端（延迟加载）
        self.llm: Optional[BaseLLMClient] = None

        logger.info(f"摘要生成器初始化: provider={self.provider}, model={self.model}")

    def _init_llm(self):
        """延迟初始化LLM"""
        if self.llm is None:
            self.llm = get_llm_client(
                provider=self.provider,
                model=self.model
            )

    def summarize(
        self,
        doc_name: str,
        content: str,
        regulation_type: str = "",
        max_tokens: Optional[int] = None
    ) -> DocumentSummary:
        """
        生成文档摘要

        Args:
            doc_name: 文档名称
            content: 文档内容（Markdown格式）
            regulation_type: 法规类型
            max_tokens: 最大输出token数

        Returns:
            DocumentSummary: 摘要结果
        """
        import time
        start_time = time.time()

        logger.info(f"生成文档摘要: {doc_name}")

        try:
            self._init_llm()

            # 使用摘要模板
            template = get_prompt_template("document_summary")

            # 构建用户消息
            user_content = template.user_template.format(
                doc_name=doc_name,
                content=content[:8000]  # 截取前8000字符（避免超出token限制）
            )

            # 调用LLM
            response = self.llm.chat(
                messages=[
                    {"role": "system", "content": template.system_prompt},
                    {"role": "user", "content": user_content}
                ],
                max_tokens=max_tokens or self.max_tokens,
                temperature=0.3  # 低温度保证摘要准确性
            )

            latency_ms = int((time.time() - start_time) * 1000)

            if not response.is_success:
                return DocumentSummary(
                    doc_name=doc_name,
                    summary="",
                    applicable_scope="",
                    key_clauses=[],
                    key_terms=[],
                    compliance_points=[],
                    model=self.model,
                    latency_ms=latency_ms,
                    error=response.error
                )

            # 解析摘要结构
            summary_data = self._parse_summary(response.content)

            logger.success(f"摘要生成完成: {doc_name}, {latency_ms}ms")

            return DocumentSummary(
                doc_name=doc_name,
                summary=summary_data.get("summary", response.content),
                applicable_scope=summary_data.get("applicable_scope", ""),
                key_clauses=summary_data.get("key_clauses", []),
                key_terms=summary_data.get("key_terms", []),
                compliance_points=summary_data.get("compliance_points", []),
                model=response.model,
                latency_ms=latency_ms
            )

        except Exception as e:
            logger.error(f"摘要生成失败: {e}")
            return DocumentSummary(
                doc_name=doc_name,
                summary="",
                applicable_scope="",
                key_clauses=[],
                key_terms=[],
                compliance_points=[],
                model=self.model,
                latency_ms=0,
                error=str(e)
            )

    def _parse_summary(self, content: str) -> Dict:
        """解析摘要内容（提取结构化信息）"""
        result = {
            "summary": content,
            "applicable_scope": "",
            "key_clauses": [],
            "key_terms": [],
            "compliance_points": []
        }

        # 简单解析（提取关键信息）
        lines = content.split("\n")

        for line in lines:
            line = line.strip()

            # 提取适用范围
            if "适用范围" in line or "适用对象" in line:
                result["applicable_scope"] = line.split("：")[-1].strip() if "：" in line else line.split(":")[-1].strip()

            # 提取关键条款
            if line.startswith("- 【条款") or line.startswith("【条款"):
                result["key_clauses"].append(line)

            # 提取关键术语
            if "关键术语" in line or "术语定义" in line:
                # 继续读取后续几行
                pass

            # 提取合规要点
            if "合规要点" in line or "必须满足" in line:
                pass

        return result

    def batch_summarize(
        self,
        documents: list
    ) -> list:
        """
        批量生成摘要

        Args:
            documents: 文档列表 [{"doc_name": str, "content": str}, ...]

        Returns:
            list: 摘要结果列表
        """
        results = []
        for doc in documents:
            result = self.summarize(doc["doc_name"], doc["content"])
            results.append(result)
        return results


def summarize_document(
    doc_name: str,
    content: str,
    **kwargs
) -> DocumentSummary:
    """便捷函数：生成文档摘要"""
    summarizer = DocumentSummarizer(**kwargs)
    return summarizer.summarize(doc_name, content)