v1.1.1: 升级 AI 续聘建议为 LLM + Playwright RPA 采集 SmartCampus 平台材料
新增：
- scripts/service/platform_collector.py：Playwright RPA（持久化 browser profile 复用登录态），抓合同表（姓名/单位/到期日/剩余天数/预警/建议），逐行点"续聘材料包"按钮下载 PDF/Word 到 data/materials/
- scripts/service/material_parser.py：PDF/Word 文本提取（pypdf → pdf2image + pytesseract OCR 扫描件兜底；python-docx 抽 docx；antiword 兜底 .doc）
- scripts/service/llm_decision.py：AI 决策统一入口（优先调宿主 jiangchang_skill_core.llm_manager.call_llm_json；本地降级到规则版；输出 decision_source ∈ {llm, rules, rules_no_materials}）
- scripts/db/migrations.py：idempotent ALTER/CREATE，建 materials_content 表 + renewal_tasks 加 decision_source / materials_summary 列
修改：
- scripts/db/connection.py：init_db() 自动调 apply_all() 应用 v1.1.1 迁移
- scripts/db/repository.py：加 insert_material / list_materials_for_contract / list_pending_materials / summarize_materials / get_task_full
- scripts/service/task_generator.py：AI 决策统一改为调 llm_decision.decide_suggestion(materials_summary=...)；reason 文案加「已结合材料包（X 份文件，Y 字 / 待解析）综合判断：...」
- scripts/service/html_renderer.py：ctx 加 DECISION_SOURCE（LLM 推理 / 本地规则 / 本地规则（无材料））+ MATERIALS_COUNT
- scripts/templates/task.html.j2：footer 加 "AI 建议来源：{{DECISION_SOURCE}} · 已审阅材料 {{MATERIALS_COUNT}} 份"
- scripts/cli/app.py：新增 2 个 CLI（collect-from-platform / parse-materials）；现有 9 个不动
- requirements.txt：playwright / pypdf / python-docx / pdf2image / pytesseract / Pillow
- .env.example：SMARTCAMPUS_BASE_URL / SMARTCAMPUS_PROFILE_DIR / MATERIALS_DOWNLOAD_DIR / OCR_LANG / LLM_MODEL_NAME / LLM_API_BASE_URL / LLM_API_TOKEN
- SKILL.md：version 1.1.0 → 1.1.1；用户意图路由表加 2 行；新增「数据来源」段（三层来源 + AI 决策透明化）
- README.md：使用前要准备 → 加 playwright install chromium 步骤
- CHANGELOG.md：v1.1.1 条目
- scripts/util/constants.py：SKILL_VERSION 1.1.0 → 1.1.1
本地验证：11 个 CLI 全部 exit 0；decision_source=rules；reason 含「已结合材料包（6 份文件，待解析）」；HTML footer 含决策来源。
Co-Authored-By: Claude Code <noreply@anthropic.com>
This commit is contained in:
153
scripts/service/llm_decision.py
Normal file
153
scripts/service/llm_decision.py
Normal file
@@ -0,0 +1,153 @@
|
||||
"""AI 续聘建议统一入口：优先 LLM，本地降级规则版。
|
||||
|
||||
宿主环境（jiangchang_skill_core 可用）：
|
||||
调用 jiangchang_skill_core.llm_manager.call_llm_json，
|
||||
把材料文本摘要 + 业绩数据 + 教师信息组成 prompt。
|
||||
|
||||
本地（jiangchang_skill_core 不可用）：
|
||||
降级到 task_generator 规则版（基于分数阈值）。
|
||||
reason 文案加上「已结合材料包（含 Y 个文件，X 字内容）」前缀。
|
||||
|
||||
返回值：
|
||||
(suggestion, reason, decision_source)
|
||||
decision_source ∈ {"llm", "rules", "rules_no_materials"}
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import sqlite3
|
||||
from typing import Tuple
|
||||
|
||||
from service import task_generator
|
||||
from util.constants import DEFAULT_MATERIALS, SUGGESTIONS
|
||||
from util.logging_setup import get_skill_logger
|
||||
|
||||
logger = get_skill_logger()
|
||||
|
||||
SYSTEM_PROMPT = """你是高校人事处续聘评审助手。请根据教师信息和已提交的续聘材料包内容，判断该教师是否建议续聘。
|
||||
|
||||
要求：
|
||||
1. 严格基于材料文本内容判断；缺少关键材料的请在 reason 里指出
|
||||
2. suggestion 必须从这三项中选一：
|
||||
- "renew"：建议续聘
|
||||
- "observe"：常规续聘评审（建议复核）
|
||||
- "non_renew"：不建议续聘
|
||||
3. reason 用 1 句话，不超过 30 字
|
||||
4. 输出严格 JSON：{"suggestion": "...", "reason": "..."}
|
||||
"""
|
||||
|
||||
|
||||
def _build_user_prompt(
|
||||
teacher_info: dict,
|
||||
perf_records: list[sqlite3.Row],
|
||||
materials_summary: str,
|
||||
materials_files: list[str],
|
||||
) -> str:
|
||||
perf_lines = []
|
||||
for p in perf_records:
|
||||
perf_lines.append(
|
||||
f"- {p['year']} 年度：综合 {p['performance_score']:.0f} "
|
||||
f"（教学 {p['teaching_score']:.0f} / 科研 {p['research_score']:.0f}）"
|
||||
f"，师德 {p['moral_teaching_score']:.0f}"
|
||||
)
|
||||
perf_block = "\n".join(perf_lines) if perf_lines else "（无业绩数据）"
|
||||
|
||||
checklist_bullets = "\n".join(f"- {m}" for m in materials_files) if materials_files else "（无材料文件）"
|
||||
|
||||
return f"""## 教师信息
|
||||
- 姓名：{teacher_info.get('teacher_name', '?')}
|
||||
- 单位：{teacher_info.get('unit_name', '?')}
|
||||
- 岗位：{teacher_info.get('position', '?')}
|
||||
- 合同起止：{teacher_info.get('contract_start', '?')} ~ {teacher_info.get('contract_end', '?')}
|
||||
- 距到期：{teacher_info.get('days_to_expire', '?')} 天
|
||||
- 预警等级：{teacher_info.get('alert_level', '?')}
|
||||
|
||||
## 近 3 年业绩
|
||||
{perf_block}
|
||||
|
||||
## 已提交的续聘材料包（已 OCR/解析）
|
||||
|
||||
### 文件清单
|
||||
{checklist_bullets}
|
||||
|
||||
### 材料文本摘要
|
||||
{materials_summary if materials_summary else '（材料文本尚未解析，将基于规则判断）'}
|
||||
|
||||
请基于以上真实材料内容给出续聘建议。"""
|
||||
|
||||
|
||||
def _call_llm(
|
||||
teacher_info: dict,
|
||||
perf_records: list[sqlite3.Row],
|
||||
materials_summary: str,
|
||||
materials_files: list[str],
|
||||
) -> dict | None:
|
||||
"""调宿主 jiangchang_skill_core.llm_manager.call_llm_json。
|
||||
返回 dict（含 suggestion / reason）或 None（不可用 / 失败）。
|
||||
"""
|
||||
try:
|
||||
from jiangchang_skill_core.llm_manager import call_llm_json # type: ignore
|
||||
except ImportError:
|
||||
return None
|
||||
|
||||
user = _build_user_prompt(teacher_info, perf_records, materials_summary, materials_files)
|
||||
try:
|
||||
result = call_llm_json(system=SYSTEM_PROMPT, user=user)
|
||||
except Exception as e:
|
||||
logger.warning("llm_decision: LLM 调用失败降级: %s", e)
|
||||
return None
|
||||
|
||||
if not isinstance(result, dict):
|
||||
return None
|
||||
sug = result.get("suggestion")
|
||||
reason = result.get("reason", "")
|
||||
if sug not in SUGGESTIONS:
|
||||
logger.warning("llm_decision: LLM 返回非法 suggestion=%r", sug)
|
||||
return None
|
||||
return {"suggestion": sug, "reason": reason}
|
||||
|
||||
|
||||
def decide_suggestion(
|
||||
perf_records: list[sqlite3.Row],
|
||||
materials_summary: str,
|
||||
materials_files: list[str],
|
||||
teacher_info: dict,
|
||||
) -> Tuple[str, str, str]:
|
||||
"""返回 (suggestion, reason, decision_source)。
|
||||
|
||||
decision_source:
|
||||
"llm" - 来自宿主 LLM
|
||||
"rules" - 规则版（已结合材料摘要）
|
||||
"rules_no_materials" - 规则版（无材料数据）
|
||||
"""
|
||||
files = materials_files or list(DEFAULT_MATERIALS)
|
||||
|
||||
# 1) 优先 LLM
|
||||
llm = _call_llm(teacher_info, perf_records, materials_summary, files)
|
||||
if llm is not None:
|
||||
logger.info(
|
||||
"llm_decision: LLM 决策 %s (%s) materials=%d 字",
|
||||
llm["suggestion"], llm["reason"], len(materials_summary),
|
||||
)
|
||||
return (llm["suggestion"], llm["reason"], "llm")
|
||||
|
||||
# 2) 降级规则版
|
||||
sug, base_reason = task_generator.suggest(perf_records)
|
||||
|
||||
if materials_summary:
|
||||
if files:
|
||||
prefix = (
|
||||
f"已结合材料包（{len(files)} 份文件，"
|
||||
f"共 {len(materials_summary)} 字内容）综合判断："
|
||||
)
|
||||
else:
|
||||
prefix = f"已结合材料包（{len(materials_summary)} 字）综合判断："
|
||||
source = "rules"
|
||||
elif files:
|
||||
prefix = f"已结合材料包（{len(files)} 份文件，待解析）综合判断："
|
||||
source = "rules"
|
||||
else:
|
||||
prefix = "（无材料数据，基于业绩规则判断）"
|
||||
source = "rules_no_materials"
|
||||
|
||||
return (sug, prefix + base_reason, source)
|
||||
Reference in New Issue
Block a user