292 lines
9.1 KiB
Python
292 lines
9.1 KiB
Python
"""
|
||
services/reference_service.py
|
||
参考范文加载服务：报告生成时按需加载对应章节参考范文
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
import logging
|
||
import re
|
||
from typing import Optional
|
||
|
||
from sqlalchemy.orm import Session
|
||
|
||
from database.models import ReportSectionReference
|
||
from services.llm_client import chat_completions_json
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
|
||
_DESENSITIZE_SYSTEM_PROMPT = """你是一个文档脱敏助手。你的任务是对后评价报告范文进行脱敏处理，只保留报告的结构骨架。
|
||
|
||
## 脱敏规则
|
||
|
||
### 必须保留的结构
|
||
1. Markdown 标题层级（## 1.1、## 1.2、### 1.2.1 等）
|
||
2. 表格的表头行、分隔行（|--|--|）
|
||
3. 段落/章节的组织顺序和逻辑关系
|
||
4. 文字的叙述逻辑（先写什么、再写什么）
|
||
5. 表格的行数、列数、表头字段名（如"序号""项目名称""可研报告""实际值"）
|
||
|
||
### 必须替换为 xxx 的内容
|
||
1. 所有具体数字：金额、年份、百分比、数量、面积、产能、投资额等
|
||
2. 项目名称、公司名称、单位名称等专有名词（书名号/引号内的内容）
|
||
3. 表格中的数据单元格内容（保留表头）
|
||
4. 具体的日期、时间节点
|
||
5. 财务指标的具体数值（IRR、NPV、回收期等）
|
||
|
||
### 特别注意
|
||
- 不要随意增删段落或改变段落顺序
|
||
- 不要删除整个表格，只替换表格中的数据单元格
|
||
- 保持原 Markdown 格式不变
|
||
- "待补充"、"详见附表"等 固定用语 不脱敏
|
||
- 书名号《》中的内容如果是不知名的规范/标准名称（如《石油化工标准》），保留书名号但内容替换为 xxx"""
|
||
|
||
|
||
_DESENSITIZE_USER_PROMPT_TEMPLATE = """请对以下后评价报告章节进行脱敏处理，只保留结构骨架，所有具体数据替换为 xxx：
|
||
|
||
```
|
||
{content}
|
||
```
|
||
|
||
请严格按照脱敏规则处理，直接输出脱敏后的完整 Markdown 内容，不要输出任何额外说明。"""
|
||
|
||
|
||
def _desensitize_via_llm(content: str) -> str:
|
||
"""
|
||
调用大模型对参考范文进行脱敏处理。
|
||
传入完整内容，返回仅保留结构骨架、具体数据替换为 xxx 的 Markdown。
|
||
|
||
若 LLM 调用失败，退回原始内容（不脱敏优于拒绝服务）。
|
||
"""
|
||
if not content or not content.strip():
|
||
return content
|
||
|
||
user_prompt = _DESENSITIZE_USER_PROMPT_TEMPLATE.format(content=content[:12000])
|
||
|
||
logger.info("参考范文脱敏 start | content_len=%s", len(content))
|
||
|
||
try:
|
||
result = chat_completions_json(
|
||
system_prompt=_DESENSITIZE_SYSTEM_PROMPT,
|
||
user_prompt=user_prompt,
|
||
temperature=0.0,
|
||
max_tokens=16384,
|
||
timeout_sec=120,
|
||
)
|
||
raw = result.get("content") or ""
|
||
if isinstance(raw, str) and raw.strip():
|
||
# 去掉可能的 ```markdown / ``` 包裹
|
||
cleaned = re.sub(r"^```(?:markdown)?\s*", "", raw.strip(), flags=re.IGNORECASE)
|
||
cleaned = re.sub(r"\s*```$", "", cleaned)
|
||
logger.info("参考范文脱敏 done | original_len=%s | desensitized_len=%s", len(content), len(cleaned))
|
||
return cleaned.strip()
|
||
except Exception as e:
|
||
logger.warning("LLM 脱敏失败，退回原文: %s", e)
|
||
|
||
return content
|
||
|
||
|
||
def load_section_reference(
|
||
db: Session,
|
||
section_key: str,
|
||
source_file: Optional[str] = None,
|
||
*,
|
||
max_chars: int = 8000,
|
||
) -> str:
|
||
"""
|
||
加载指定章节的参考范文内容。
|
||
|
||
Args:
|
||
db: 数据库会话
|
||
section_key: 章节标识（如 "1.1", "2.1.1"）
|
||
source_file: 来源文件名（可选，不指定时取该章节最新的一条）
|
||
max_chars: 最大字符数，超出截断
|
||
|
||
Returns:
|
||
参考范文 Markdown 文本，未找到时返回空字符串
|
||
"""
|
||
query = db.query(ReportSectionReference).filter(
|
||
ReportSectionReference.section_key == section_key
|
||
)
|
||
|
||
if source_file:
|
||
query = query.filter(ReportSectionReference.source_file == source_file)
|
||
|
||
ref = (
|
||
query
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
|
||
if not ref or not ref.content:
|
||
return ""
|
||
|
||
content = ref.content.strip()
|
||
if not content:
|
||
return ""
|
||
|
||
content = _desensitize_via_llm(content)
|
||
|
||
if len(content) > max_chars:
|
||
logger.info("参考范文 %s 超出 %d 字符限制，已截断", section_key, max_chars)
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
|
||
return content
|
||
|
||
|
||
def load_section_reference_by_title(
|
||
db: Session,
|
||
section_title: str,
|
||
source_file: Optional[str] = None,
|
||
*,
|
||
max_chars: int = 8000,
|
||
) -> str:
|
||
"""
|
||
按标题关键字匹配加载参考范文（不精确匹配 section_key 时的兜底方案）。
|
||
"""
|
||
refs = db.query(ReportSectionReference)
|
||
|
||
if source_file:
|
||
refs = refs.filter(ReportSectionReference.source_file == source_file)
|
||
|
||
# 尝试精确匹配 section_key（从标题中提取编号）
|
||
import re
|
||
|
||
m = re.match(r"(\d+(?:\.\d+)*)", section_title.strip())
|
||
if m:
|
||
key = m.group(1)
|
||
exact = (
|
||
refs.filter(ReportSectionReference.section_key == key)
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
if exact and exact.content:
|
||
content = exact.content.strip()
|
||
content = _desensitize_via_llm(content)
|
||
if len(content) > max_chars:
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
return content
|
||
|
||
# 按标题模糊匹配
|
||
ref = (
|
||
refs.filter(ReportSectionReference.section_title.contains(section_title[:20]))
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
|
||
if not ref or not ref.content:
|
||
return ""
|
||
|
||
content = ref.content.strip()
|
||
if not content:
|
||
return ""
|
||
|
||
content = _desensitize_via_llm(content)
|
||
|
||
if len(content) > max_chars:
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
|
||
return content
|
||
|
||
|
||
def load_section_reference_raw(
|
||
db: Session,
|
||
section_key: str,
|
||
template_id: Optional[str] = None,
|
||
*,
|
||
max_chars: int = 8000,
|
||
) -> str:
|
||
"""
|
||
加载指定章节存储在数据库中的原始参考范文内容（不做 LLM 脱敏）。
|
||
|
||
与 load_section_reference 的区别：直接返回 report_section_references.content 原文，
|
||
仅保留长度截断保护，不再调用 _desensitize_via_llm。
|
||
|
||
template_id: 选中模板的 ID。传入后只注入与该模板关联的参考范文，实现“按模板过滤”；
|
||
为空则不做模板过滤（取最新一条）。
|
||
"""
|
||
query = db.query(ReportSectionReference).filter(
|
||
ReportSectionReference.section_key == section_key
|
||
)
|
||
|
||
if template_id:
|
||
query = query.filter(ReportSectionReference.template_id == template_id)
|
||
|
||
ref = (
|
||
query
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
|
||
if not ref or not ref.content:
|
||
return ""
|
||
|
||
content = ref.content.strip()
|
||
if not content:
|
||
return ""
|
||
|
||
if len(content) > max_chars:
|
||
logger.info("参考范文 %s 超出 %d 字符限制，已截断", section_key, max_chars)
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
|
||
return content
|
||
|
||
|
||
def load_section_reference_raw_by_title(
|
||
db: Session,
|
||
section_title: str,
|
||
template_id: Optional[str] = None,
|
||
*,
|
||
max_chars: int = 8000,
|
||
) -> str:
|
||
"""按标题匹配加载原始参考范文内容（不做 LLM 脱敏），用于 section_key 未命中时的兜底。"""
|
||
refs = db.query(ReportSectionReference)
|
||
|
||
if template_id:
|
||
refs = refs.filter(ReportSectionReference.template_id == template_id)
|
||
|
||
import re
|
||
|
||
m = re.match(r"(\d+(?:\.\d+)*)", section_title.strip())
|
||
if m:
|
||
key = m.group(1)
|
||
exact = (
|
||
refs.filter(ReportSectionReference.section_key == key)
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
if exact and exact.content:
|
||
content = exact.content.strip()
|
||
if len(content) > max_chars:
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
return content
|
||
|
||
ref = (
|
||
refs.filter(ReportSectionReference.section_title.contains(section_title[:20]))
|
||
.order_by(ReportSectionReference.updated_at.desc())
|
||
.first()
|
||
)
|
||
|
||
if not ref or not ref.content:
|
||
return ""
|
||
|
||
content = ref.content.strip()
|
||
if not content:
|
||
return ""
|
||
|
||
if len(content) > max_chars:
|
||
content = content[:max_chars] + "\n\n（参考范文超出长度限制，已截断）"
|
||
|
||
return content
|
||
|
||
|
||
def list_available_source_files(db: Session) -> list[str]:
|
||
"""列出所有已上传的参考范文来源文件列表。"""
|
||
results = (
|
||
db.query(ReportSectionReference.source_file)
|
||
.distinct()
|
||
.order_by(ReportSectionReference.source_file)
|
||
.all()
|
||
)
|
||
return [r[0] for r in results if r[0]] |