complete_agent_flow_rule.py 33 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753
  1. """
  2. 完整的智能体工作流 (Complete Agent Flow)
  3. =====================================
  4. 此工作流整合了规划、大纲生成和指标计算四个核心智能体,实现完整的报告生成流程。
  5. 包含的智能体:
  6. 1. PlanningAgent (规划智能体) - 分析状态并做出决策
  7. 2. OutlineAgent (大纲生成智能体) - 生成报告结构和指标需求
  8. 3. MetricCalculationAgent (指标计算智能体) - 执行标准指标计算
  9. 4. RulesEngineMetricCalculationAgent (规则引擎指标计算智能体) - 执行规则引擎指标计算
  10. 工作流程:
  11. 1. 规划节点 → 分析当前状态,决定下一步行动
  12. 2. 大纲生成节点 → 生成报告大纲和指标需求
  13. 3. 指标判断节点 → 根据大纲确定需要计算的指标
  14. 4. 指标计算节点 → 执行具体的指标计算任务
  15. 技术特点:
  16. - 基于LangGraph的状态机工作流
  17. - 支持条件路由和状态管理
  18. - 完善的错误处理机制
  19. - 详细的执行日志记录
  20. 作者: Big Agent Team
  21. 版本: 1.0.0
  22. 创建时间: 2024-12-20
  23. """
  24. import asyncio
  25. from typing import Dict, Any, List
  26. from datetime import datetime
  27. from langgraph.graph import StateGraph, START, END
  28. from llmops.workflow_state import (
  29. IntegratedWorkflowState,
  30. create_initial_integrated_state,
  31. get_calculation_progress,
  32. update_state_with_outline_generation,
  33. update_state_with_planning_decision,
  34. update_state_with_data_classified,
  35. convert_numpy_types,
  36. update_state_with_data_standardize
  37. )
  38. from llmops.agents.outline_agent import generate_report_outline
  39. from llmops.agents.planning_agent import plan_next_action
  40. from llmops.agents.rules_engine_metric_calculation_agent import RulesEngineMetricCalculationAgent
  41. from llmops.agents.data_manager import DataManager
  42. import os
  43. from llmops.agents.data_classify_agent import data_classify
  44. from llmops.config import DEEPSEEK_API_KEY, multimodal_api_url, LLM_API_KEY, LLM_BASE_URL, LLM_MODEL_NAME
  45. from llmops.agents.data_stardard import data_standardize
  46. class CompleteAgentFlow:
  47. """完整的智能体工作流"""
  48. def __init__(self, api_key: str, base_url: str = "https://api.deepseek.com", model_name: str = "deepseek-chat"):
  49. """
  50. 初始化完整的工作流
  51. Args:
  52. api_key: DeepSeek API密钥
  53. base_url: DeepSeek API基础URL
  54. model_name: 模型名称
  55. """
  56. self.api_key = api_key
  57. self.base_url = base_url
  58. self.model_name = model_name
  59. # 初始规则引擎智能体
  60. self.rules_engine_agent = RulesEngineMetricCalculationAgent(api_key, base_url)
  61. # 创建工作流图
  62. self.workflow = self._create_workflow()
  63. def _create_workflow(self) -> StateGraph:
  64. """创建LangGraph工作流"""
  65. workflow = StateGraph(IntegratedWorkflowState)
  66. # 添加节点
  67. workflow.add_node("planning_node", self._planning_node)
  68. workflow.add_node("outline_generator", self._outline_generator_node)
  69. workflow.add_node("metric_calculator", self._metric_calculator_node)
  70. workflow.add_node("data_classify", self._data_classify_node)
  71. workflow.add_node("data_standardize", self._data_standardize_node)
  72. # 设置入口点
  73. workflow.set_entry_point("planning_node")
  74. # 添加条件边 - 基于规划决策路由
  75. workflow.add_conditional_edges(
  76. "planning_node",
  77. self._route_from_planning,
  78. {
  79. "outline_generator": "outline_generator",
  80. "metric_calculator": "metric_calculator",
  81. "data_classify": "data_classify",
  82. "data_standardize": "data_standardize",
  83. END: END
  84. }
  85. )
  86. # 从各个节点返回规划节点重新决策
  87. workflow.add_edge("data_standardize", "planning_node")
  88. workflow.add_edge("data_classify", "planning_node")
  89. workflow.add_edge("outline_generator", "planning_node")
  90. workflow.add_edge("metric_calculator", "planning_node")
  91. return workflow
  92. def _route_from_planning(self, state: IntegratedWorkflowState) -> str:
  93. """
  94. 从规划节点路由到下一个节点
  95. Args:
  96. state: 当前状态
  97. Returns:
  98. 目标节点名称
  99. """
  100. print(f"\n🔍 [路由决策] 步骤={state['planning_step']}, "
  101. f"数据集分类打标数量={len(state.get('data_set_classified', []))}",
  102. f"大纲={state.get('outline_draft') is not None}, "
  103. f"指标需求={len(state.get('metrics_requirements', []))}")
  104. # 防止无限循环
  105. if state['planning_step'] > 30:
  106. print("⚠️ 规划步骤超过30次,强制结束流程")
  107. return END
  108. # 数据标准化状态 0 → 数据标准化
  109. if state.get("is_standardized", 0) == 0:
  110. print("→ 路由到 data_standardize(数据标准化)")
  111. return "data_standardize"
  112. # 数据分类打标数量为0 → 分类打标
  113. if len(state.get("data_set_classified", [])) == 0:
  114. print("→ 路由到 data_classify(分类打标)")
  115. return "data_classify"
  116. # 如果大纲为空 → 生成大纲
  117. if not state.get("outline_draft"):
  118. print("→ 路由到 outline_generator(生成大纲)")
  119. return "outline_generator"
  120. # 如果指标需求为空但大纲已生成 → 评估指标需求
  121. if not state.get("metrics_requirements") and state.get("outline_draft"):
  122. print("→ 路由到 metric_evaluator(评估指标需求)")
  123. return "metric_evaluator"
  124. # 计算覆盖率
  125. progress = get_calculation_progress(state)
  126. coverage = progress["coverage_rate"]
  127. print(f" 指标覆盖率 = {coverage:.2%}")
  128. # 如果有待计算指标且覆盖率 < 100% → 计算指标
  129. if state.get("pending_metric_ids") and coverage < 1.0:
  130. print(f"→ 路由到 metric_calculator(计算指标,覆盖率={coverage:.2%})")
  131. return "metric_calculator"
  132. # 检查是否应该结束流程
  133. pending_ids = state.get("pending_metric_ids", [])
  134. failed_attempts = state.get("failed_metric_attempts", {})
  135. max_retries = 3
  136. # 计算还有哪些指标可以重试(未达到最大重试次数)
  137. retryable_metrics = [
  138. mid for mid in pending_ids
  139. if failed_attempts.get(mid, 0) < max_retries
  140. ]
  141. # 如果覆盖率 >= 80%,或者没有可重试的指标 → 结束流程
  142. if coverage >= 0.8 or not retryable_metrics:
  143. reason = "覆盖率达到80%" if coverage >= 0.8 else "没有可重试指标"
  144. print(f"→ 结束流程(覆盖率={coverage:.2%},原因:{reason})")
  145. return END
  146. # 默认返回规划节点
  147. return "planning_node"
  148. async def _planning_node(self, state: IntegratedWorkflowState) -> IntegratedWorkflowState:
  149. """规划节点:分析状态并做出决策"""
  150. try:
  151. print("🧠 正在执行规划分析...")
  152. # 使用规划智能体做出决策
  153. decision = await plan_next_action(
  154. question=state["question"],
  155. industry=state["industry"],
  156. current_state=state,
  157. api_key=self.api_key,
  158. base_url=self.base_url,
  159. model_name=self.model_name
  160. )
  161. # 更新状态
  162. new_state = update_state_with_planning_decision(state, {
  163. "decision": decision.decision,
  164. "next_route": self._decision_to_route(decision.decision),
  165. "metrics_to_compute": decision.metrics_to_compute
  166. })
  167. # 添加决策消息
  168. decision_msg = self._format_decision_message(decision)
  169. new_state["messages"].append({
  170. "role": "assistant",
  171. "content": decision_msg,
  172. "timestamp": datetime.now().isoformat()
  173. })
  174. print(f"✅ 规划决策完成:{decision.decision}")
  175. return convert_numpy_types(new_state)
  176. except Exception as e:
  177. print(f"❌ 规划节点执行失败: {e}")
  178. new_state = state.copy()
  179. new_state["errors"].append(f"规划节点错误: {str(e)}")
  180. return convert_numpy_types(new_state)
  181. async def _outline_generator_node(self, state: IntegratedWorkflowState) -> IntegratedWorkflowState:
  182. """大纲生成节点"""
  183. try:
  184. print("📝 正在生成报告大纲...")
  185. # 生成大纲(支持重试机制)
  186. outline = await generate_report_outline(
  187. question=state["question"],
  188. industry=state["industry"],
  189. sample_data=state["data_set"][:3], # 使用前3个样本
  190. api_key=self.api_key,
  191. base_url=self.base_url,
  192. model_name=self.model_name,
  193. max_retries=3, # 最多重试5次
  194. retry_delay=3.0 # 每次重试间隔3秒
  195. )
  196. # 更新状态
  197. new_state = update_state_with_outline_generation(state, outline)
  198. print(f"✅ 大纲生成完成:{outline.report_title}")
  199. print(f" 包含 {len(outline.sections)} 个章节,{len(outline.global_metrics)} 个指标需求")
  200. # 分析并打印AI的指标选择推理过程
  201. self._print_ai_selection_analysis(outline)
  202. return convert_numpy_types(new_state)
  203. except Exception as e:
  204. print(f"❌ 大纲生成失败: {e}")
  205. new_state = state.copy()
  206. new_state["errors"].append(f"大纲生成错误: {str(e)}")
  207. return convert_numpy_types(new_state)
  208. async def _data_classify_node(self, state: IntegratedWorkflowState) -> IntegratedWorkflowState:
  209. """数据分类打标节点"""
  210. try:
  211. standardized_file_path = state["standardized_file_path"]
  212. file_name = os.path.basename(standardized_file_path)
  213. # 读取标准化后的数据文件
  214. data_set = DataManager.load_data_from_csv_file(standardized_file_path)
  215. # 加载测试数据集并展示两条样例
  216. print(f"📊 读取标准化数据文件: {file_name}, 加载 {len(data_set)} 条记录")
  217. print(f"测试数据样例: {data_set[0:1]}")
  218. print("📝 正在对数据进行分类打标...")
  219. # 对数据进行分类打标
  220. data_set_classified = await data_classify(
  221. industry=state["industry"],
  222. data_set=data_set,
  223. file_name=state["file_name"]
  224. )
  225. # 更新状态
  226. new_state = update_state_with_data_classified(state, data_set_classified)
  227. print(f"✅ 数据分类打标完成,打标记录数: {len(data_set_classified)}")
  228. return convert_numpy_types(new_state)
  229. except Exception as e:
  230. print(f"❌ 数据分类打标失败: {e}")
  231. new_state = state.copy()
  232. new_state["errors"].append(f"数据分类打标错误: {str(e)}")
  233. return convert_numpy_types(new_state)
  234. async def _data_standardize_node(self, state: IntegratedWorkflowState) -> IntegratedWorkflowState:
  235. """数据标准化节点"""
  236. try:
  237. print("📝 正在对数据进行标准化处理...")
  238. # 数据标准化处理
  239. result = await data_standardize(
  240. api_key=self.api_key,
  241. base_url=self.base_url,
  242. model_name=self.model_name,
  243. multimodal_api_url=multimodal_api_url,
  244. input_file_path=state["original_file_path"]
  245. )
  246. is_succ = 0
  247. standardized_file_path = None
  248. if result["status"] == "success": # 数据标准化成功
  249. is_succ = 1
  250. standardized_file_path = result["file_path"]
  251. # 更新状态
  252. new_state = update_state_with_data_standardize(state, is_succ, standardized_file_path)
  253. print(f"✅ 数据标准化完成,处理状态: {is_succ},标准化文件路径:{standardized_file_path}")
  254. return convert_numpy_types(new_state)
  255. except Exception as e:
  256. print(f"❌ 数据标准化失败: {e}")
  257. new_state = state.copy()
  258. new_state["errors"].append(f"数据标准化错误: {str(e)}")
  259. return convert_numpy_types(new_state)
  260. def _print_ai_selection_analysis(self, outline):
  261. """打印AI指标选择的推理过程分析 - 完全通用版本"""
  262. print()
  263. print('╔══════════════════════════════════════════════════════════════════════════════╗')
  264. print('║ 🤖 AI指标选择分析 ║')
  265. print('╚══════════════════════════════════════════════════════════════════════════════╝')
  266. print()
  267. # 计算总指标数 - outline可能是字典格式,需要适配
  268. if hasattr(outline, 'sections'):
  269. # Pydantic模型格式
  270. total_metrics = sum(len(section.metrics_needed) for section in outline.sections)
  271. sections = outline.sections
  272. else:
  273. # 字典格式
  274. total_metrics = sum(len(section.get('metrics_needed', [])) for section in outline.get('sections', []))
  275. sections = outline.get('sections', [])
  276. # 获取可用指标总数(这里可以从状态或其他地方动态获取)
  277. available_count = 26 # 这个可以从API调用中动态获取
  278. print('📊 选择统计:')
  279. print(' ┌─────────────────────────────────────────────────────────────────────┐')
  280. print(' │ 系统可用指标: {}个 │ AI本次选择: {}个 │ 选择率: {:.1f}% │'.format(
  281. available_count, total_metrics, total_metrics/available_count*100 if available_count > 0 else 0))
  282. print(' └─────────────────────────────────────────────────────────────────────┘')
  283. print()
  284. print('📋 AI决策过程:')
  285. print(' 大模型已根据用户需求从{}个可用指标中选择了{}个最相关的指标。'.format(available_count, total_metrics))
  286. print(' 选择过程完全由大模型基于语义理解和业务逻辑进行,不涉及任何硬编码规则。')
  287. print()
  288. print('🔍 选择结果:')
  289. print(' • 总章节数: {}个'.format(len(sections)))
  290. print(' • 平均每章节指标数: {:.1f}个'.format(total_metrics/len(sections) if sections else 0))
  291. print(' • 选择策略: 基于用户需求的相关性分析')
  292. print()
  293. print('🎯 AI Agent核心能力:')
  294. print(' • 语义理解: 理解用户查询的业务意图和分析需求')
  295. print(' • 智能筛选: 从海量指标中挑选最相关的组合')
  296. print(' • 逻辑推理: 为每个分析维度提供充分的选择依据')
  297. print(' • 动态适配: 根据不同场景自动调整选择策略')
  298. print()
  299. print('💡 关键洞察:')
  300. print(' AI Agent通过大模型的推理能力,实现了超越传统规则引擎的智能化指标选择,')
  301. print(' 能够根据具体业务场景动态调整分析框架,确保分析的针对性和有效性。')
  302. print()
  303. async def _metric_calculator_node(self, state: IntegratedWorkflowState) -> IntegratedWorkflowState:
  304. """指标计算节点"""
  305. try:
  306. # 检查计算模式
  307. use_rules_engine_only = state.get("use_rules_engine_only", False)
  308. use_traditional_engine_only = state.get("use_traditional_engine_only", False)
  309. if use_rules_engine_only:
  310. print("🧮 正在执行规则引擎指标计算(专用模式)...")
  311. elif use_traditional_engine_only:
  312. print("🧮 正在执行传统引擎指标计算(专用模式)...")
  313. else:
  314. print("🧮 正在执行指标计算...")
  315. new_state = state.copy()
  316. # 使用规划决策指定的指标批次,如果没有指定则使用所有待计算指标
  317. current_batch = state.get("current_batch_metrics", [])
  318. if current_batch:
  319. pending_ids = current_batch
  320. print(f"🧮 本次计算批次包含 {len(pending_ids)} 个指标")
  321. else:
  322. pending_ids = state.get("pending_metric_ids", [])
  323. print(f"🧮 计算所有待计算指标,共 {len(pending_ids)} 个")
  324. if not pending_ids:
  325. print("⚠️ 没有待计算的指标")
  326. return convert_numpy_types(new_state)
  327. # 获取指标需求信息
  328. metrics_requirements = state.get("metrics_requirements", [])
  329. if not metrics_requirements:
  330. print("⚠️ 没有指标需求信息")
  331. return convert_numpy_types(new_state)
  332. # 计算成功和失败的指标
  333. successful_calculations = 0
  334. failed_calculations = 0
  335. # 遍历待计算的指标(创建副本避免修改时遍历的问题)
  336. for metric_id in pending_ids.copy():
  337. try:
  338. # 找到对应的指标需求
  339. metric_req = next((m for m in metrics_requirements if m.metric_id == metric_id), None)
  340. if not metric_req:
  341. # 修复:找不到指标需求时,创建临时的指标需求结构,避免跳过指标
  342. print(f"⚠️ 指标 {metric_id} 找不到需求信息,创建临时配置继续计算")
  343. metric_req = type('MetricRequirement', (), {
  344. 'metric_id': metric_id,
  345. 'metric_name': metric_id.replace('metric-', '') if metric_id.startswith('metric-') else metric_id,
  346. 'calculation_logic': f'计算 {metric_id}',
  347. 'required_fields': ['transactions'],
  348. 'dependencies': []
  349. })()
  350. print(f"🧮 计算指标: {metric_id} - {metric_req.metric_name}")
  351. # 根据模式决定使用哪种计算方式
  352. if use_rules_engine_only:
  353. # 只使用规则引擎计算
  354. use_rules_engine = True
  355. print(f" 使用规则引擎模式")
  356. elif use_traditional_engine_only:
  357. # 只使用传统引擎计算
  358. use_rules_engine = False
  359. print(f" 使用传统引擎模式")
  360. else:
  361. # 自动选择计算方式:优先使用规则引擎,只在规则引擎不可用时使用传统计算
  362. use_rules_engine = True # 默认使用规则引擎计算所有指标
  363. if use_rules_engine:
  364. # 使用规则引擎计算
  365. # 现在metric_id已经是知识ID,直接使用它作为配置名
  366. config_name = metric_id # metric_id 已经是知识ID,如 "metric-分析账户数量"
  367. intent_result = {
  368. "target_configs": [config_name],
  369. "intent_category": "指标计算"
  370. }
  371. print(f" 使用知识ID: {config_name}")
  372. # 将打好标的数据集传入指标计算函数中
  373. data_set_classified = state.get("data_set_classified", [])
  374. results = await self.rules_engine_agent.calculate_metrics(intent_result, data_set_classified)
  375. else:
  376. # 使用传统指标计算(模拟)
  377. # 这里简化处理,实际应该根据配置文件调用相应的API
  378. results = {
  379. "success": True,
  380. "results": [{
  381. "config_name": metric_req.metric_id,
  382. "result": {
  383. "success": True,
  384. "data": f"传统引擎计算结果:{metric_req.metric_name}",
  385. "value": 100.0 # 模拟数值
  386. }
  387. }]
  388. }
  389. # 处理计算结果
  390. calculation_success = False
  391. for result in results.get("results", []):
  392. if result.get("result", {}).get("success"):
  393. # 计算成功
  394. new_state["computed_metrics"][metric_id] = result["result"]
  395. successful_calculations += 1
  396. calculation_success = True
  397. print(f"✅ 指标 {metric_id} 计算成功")
  398. break # 找到一个成功的就算成功
  399. else:
  400. # 计算失败
  401. failed_calculations += 1
  402. print(f"❌ 指标 {metric_id} 计算失败")
  403. # 初始化失败尝试记录
  404. if "failed_metric_attempts" not in new_state:
  405. new_state["failed_metric_attempts"] = {}
  406. # 根据计算结果处理指标
  407. if calculation_success:
  408. # 计算成功:从待计算列表中移除
  409. if metric_id in new_state["pending_metric_ids"]:
  410. new_state["pending_metric_ids"].remove(metric_id)
  411. # 重置失败计数
  412. new_state["failed_metric_attempts"].pop(metric_id, None)
  413. else:
  414. # 计算失败:记录失败次数,不从待计算列表移除
  415. new_state["failed_metric_attempts"][metric_id] = new_state["failed_metric_attempts"].get(metric_id, 0) + 1
  416. max_retries = 3
  417. if new_state["failed_metric_attempts"][metric_id] >= max_retries:
  418. print(f"⚠️ 指标 {metric_id} 已达到最大重试次数 ({max_retries}),从待计算列表中移除")
  419. if metric_id in new_state["pending_metric_ids"]:
  420. new_state["pending_metric_ids"].remove(metric_id)
  421. except Exception as e:
  422. print(f"❌ 计算指标 {metric_id} 时发生异常: {e}")
  423. failed_calculations += 1
  424. # 初始化失败尝试记录
  425. if "failed_metric_attempts" not in new_state:
  426. new_state["failed_metric_attempts"] = {}
  427. # 记录失败次数
  428. new_state["failed_metric_attempts"][metric_id] = new_state["failed_metric_attempts"].get(metric_id, 0) + 1
  429. max_retries = 3
  430. if new_state["failed_metric_attempts"][metric_id] >= max_retries:
  431. print(f"⚠️ 指标 {metric_id} 异常已达到最大重试次数 ({max_retries}),从待计算列表中移除")
  432. if metric_id in new_state["pending_metric_ids"]:
  433. new_state["pending_metric_ids"].remove(metric_id)
  434. # 更新计算结果统计
  435. new_state["calculation_results"] = {
  436. "total_configs": len(pending_ids),
  437. "successful_calculations": successful_calculations,
  438. "failed_calculations": failed_calculations
  439. }
  440. # 添加消息
  441. if use_rules_engine_only:
  442. message_content = f"🧮 规则引擎指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败"
  443. elif use_traditional_engine_only:
  444. message_content = f"🧮 传统引擎指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败"
  445. else:
  446. message_content = f"🧮 指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败"
  447. new_state["messages"].append({
  448. "role": "assistant",
  449. "content": message_content,
  450. "timestamp": datetime.now().isoformat()
  451. })
  452. if use_rules_engine_only:
  453. print(f"✅ 规则引擎指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败")
  454. elif use_traditional_engine_only:
  455. print(f"✅ 传统引擎指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败")
  456. else:
  457. print(f"✅ 指标计算完成:{successful_calculations} 成功,{failed_calculations} 失败")
  458. return convert_numpy_types(new_state)
  459. except Exception as e:
  460. print(f"❌ 指标计算节点失败: {e}")
  461. new_state = state.copy()
  462. new_state["errors"].append(f"指标计算错误: {str(e)}")
  463. return convert_numpy_types(new_state)
  464. def _decision_to_route(self, decision: str) -> str:
  465. """将规划决策转换为路由"""
  466. decision_routes = {
  467. "data_classify": "data_classify",
  468. "generate_outline": "outline_generator",
  469. "compute_metrics": "metric_calculator",
  470. "finalize_report": END # 直接结束流程
  471. }
  472. return decision_routes.get(decision, "planning_node")
  473. def _format_decision_message(self, decision: Any) -> str:
  474. """格式化决策消息"""
  475. try:
  476. decision_type = getattr(decision, 'decision', 'unknown')
  477. reasoning = getattr(decision, 'reasoning', '')
  478. if decision_type == "compute_metrics" and hasattr(decision, 'metrics_to_compute'):
  479. metrics = decision.metrics_to_compute
  480. return f"🧮 规划决策:计算 {len(metrics)} 个指标"
  481. elif decision_type == "finalize_report":
  482. return f"✅ 规划决策:生成最终报告"
  483. elif decision_type == "generate_outline":
  484. return f"📋 规划决策:生成大纲"
  485. else:
  486. return f"🤔 规划决策:{decision_type}"
  487. except:
  488. return "🤔 规划决策已完成"
  489. async def run_workflow(self, question: str, industry: str, original_file_path: str, session_id: str = None, use_rules_engine_only: bool = False, use_traditional_engine_only: bool = False) -> Dict[str, Any]:
  490. """
  491. 运行完整的工作流
  492. Args:
  493. question: 用户查询
  494. industry: 行业
  495. data: 数据集
  496. original_file_path: 原始文件路径
  497. session_id: 会话ID
  498. use_rules_engine_only: 是否只使用规则引擎指标计算
  499. use_traditional_engine_only: 是否只使用传统引擎指标计算
  500. Returns:
  501. 工作流结果
  502. """
  503. try:
  504. print("🚀 启动完整智能体工作流...")
  505. print(f"问题:{question}")
  506. print(f"行业:{industry}")
  507. print(f"数据文件:{original_file_path}")
  508. # print(f"数据条数:{len(data)}")
  509. if use_rules_engine_only:
  510. print("计算模式:只使用规则引擎")
  511. elif use_traditional_engine_only:
  512. print("计算模式:只使用传统引擎")
  513. else:
  514. print("计算模式:标准模式")
  515. # 创建初始状态
  516. initial_state = create_initial_integrated_state(question, industry, original_file_path, session_id)
  517. # 设置计算模式标记
  518. if use_rules_engine_only:
  519. initial_state["use_rules_engine_only"] = True
  520. initial_state["use_traditional_engine_only"] = False
  521. elif use_traditional_engine_only:
  522. initial_state["use_rules_engine_only"] = False
  523. initial_state["use_traditional_engine_only"] = True
  524. else:
  525. initial_state["use_rules_engine_only"] = False
  526. initial_state["use_traditional_engine_only"] = False
  527. # 编译工作流
  528. app = self.workflow.compile()
  529. # 执行工作流
  530. result = await app.ainvoke(initial_state)
  531. print("✅ 工作流执行完成")
  532. return {
  533. "success": True,
  534. "result": result,
  535. "answer": result.get("answer"),
  536. "report": result.get("report_draft"),
  537. "session_id": result.get("session_id"),
  538. "execution_summary": {
  539. "planning_steps": result.get("planning_step", 0),
  540. "outline_generated": result.get("outline_draft") is not None,
  541. "metrics_computed": len(result.get("computed_metrics", {})),
  542. "completion_rate": result.get("completeness_score", 0)
  543. }
  544. }
  545. except Exception as e:
  546. print(f"❌ 工作流执行失败: {e}")
  547. return {
  548. "success": False,
  549. "error": str(e),
  550. "result": None
  551. }
  552. # 便捷函数
  553. async def run_complete_agent_flow(question: str, industry: str, data: List[Dict[str, Any]], file_name: str, api_key: str, session_id: str = None, use_rules_engine_only: bool = False, use_traditional_engine_only: bool = False) -> Dict[str, Any]:
  554. """
  555. 运行完整智能体工作流的便捷函数
  556. Args:
  557. question: 用户查询
  558. data: 数据集
  559. file_name: 数据文件名称
  560. api_key: API密钥
  561. session_id: 会话ID
  562. use_rules_engine_only: 是否只使用规则引擎指标计算
  563. use_traditional_engine_only: 是否只使用传统引擎指标计算
  564. Returns:
  565. 工作流结果
  566. """
  567. workflow = CompleteAgentFlow(api_key)
  568. return await workflow.run_workflow(question, industry, data, file_name, session_id, use_rules_engine_only, use_traditional_engine_only)
  569. # 便捷函数
  570. async def run_flow(question: str, industry: str, original_file_path: str, api_key: str, base_url: str, model_name: str, session_id: str = None, use_rules_engine_only: bool = False, use_traditional_engine_only: bool = False) -> Dict[str, Any]:
  571. """
  572. 运行完整智能体工作流的便捷函数
  573. Args:
  574. question: 用户查询
  575. data: 数据集
  576. original_file_path: 原始文件路径(pdf/img/csv)
  577. api_key: API密钥
  578. base_url: LLM base url
  579. model_name: LLM model name
  580. session_id: 会话ID
  581. use_rules_engine_only: 是否只使用规则引擎指标计算
  582. use_traditional_engine_only: 是否只使用传统引擎指标计算
  583. Returns:
  584. 工作流结果
  585. """
  586. workflow = CompleteAgentFlow(api_key, base_url, model_name)
  587. return await workflow.run_workflow(question, industry, original_file_path, session_id, use_rules_engine_only, use_traditional_engine_only)
  588. # 主函数用于测试
  589. async def main():
  590. """主函数:执行系统测试"""
  591. import os
  592. os.environ["LANGCHAIN_TRACING_V2"] = "false"
  593. os.environ["LANGCHAIN_API_KEY"] = ""
  594. # 禁用 LangGraph 的追踪
  595. os.environ["LANGSMITH_TRACING"] = "false"
  596. print("🚀 执行CompleteAgentFlow系统测试")
  597. print("=" * 50)
  598. # 行业
  599. industry = "农业"
  600. # 测试文件(pdf/img/csv)
  601. file_name = "11111.png"
  602. curr_dir = os.path.dirname(os.path.abspath(__file__))
  603. file_path = os.path.join(curr_dir, "..", "data_files", file_name)
  604. print(f"使用LLM:{LLM_MODEL_NAME}")
  605. # 执行测试
  606. result = await run_flow(
  607. question="请生成一份详细的农业经营贷流水分析报告,需要包含:1.总收入和总支出统计 2.收入笔数和支出笔数 3.各类型收入支出占比分析 4.交易对手收入支出TOP3排名 5.按月份的收入支出趋势分析 6.账户数量和交易时间范围统计 7.资金流入流出月度统计等全面指标",
  608. industry = industry,
  609. original_file_path=file_path,
  610. api_key=LLM_API_KEY,
  611. base_url=LLM_BASE_URL,
  612. model_name=LLM_MODEL_NAME,
  613. session_id="direct-test"
  614. )
  615. print(f"📋 结果: {'✅ 成功' if result.get('success') else '❌ 失败'}")
  616. if result.get('success'):
  617. summary = result.get('execution_summary', {})
  618. print(f" 规划步骤: {summary.get('planning_steps', 0)}")
  619. print(f" 指标计算: {summary.get('metrics_computed', 0)}")
  620. print("🎉 测试成功!")
  621. return result
  622. if __name__ == "__main__":
  623. import asyncio
  624. asyncio.run(main())