# TradingAgents 输入输出检查指南 ## 📍 快速定位 ### 1. **状态定义位置** (输入/输出的数据结构) **文件**: `tradingagents/agents/utils/agent_states.py` 这是所有输入输出的数据结构定义，包含三个阶段的所有字段： ```python class AgentState: # 输入 company_of_interest: str # 公司名称 trade_date: str # 交易日期 messages: List[Message] # 消息历史 # 阶段1输出: 分析师报告 market_report: str # 市场分析师报告 sentiment_report: str # 社交媒体分析师报告 news_report: str # 新闻分析师报告 fundamentals_report: str # 基本面分析师报告 # 阶段2输出: 投资辩论 investment_debate_state: InvestDebateState # 包含 bull_history, bear_history, judge_decision investment_plan: str # 研究经理的投资计划 trader_investment_plan: str # 交易员的投资计划 # 阶段3输出: 风险分析 risk_debate_state: RiskDebateState # 包含 risky/safe/neutral 历史 final_trade_decision: str # 最终交易决策 ``` --- ## 🔍 各阶段输入输出检查位置 ### **阶段 1: 分析师阶段 (Analyst Phase)** #### 输入检查: - **初始状态**: `tradingagents/graph/propagation.py` 第 18-42 行 ```python init_state = { "company_of_interest": company_name, "trade_date": trade_date, "market_report": "", # 初始为空 "sentiment_report": "", # 初始为空 "news_report": "", # 初始为空 "fundamentals_report": "", # 初始为空 ... } ``` #### 输出检查: - **运行时 (Debug模式)**: `tradingagents/graph/trading_graph.py` 第 171-179 行 ```python for chunk in self.graph.stream(init_agent_state, **args): # chunk 包含每个节点的输出 if "market_report" in chunk: print(chunk["market_report"]) ``` - **最终状态**: `tradingagents/graph/trading_graph.py` 第 184 行 ```python final_state = self.graph.invoke(init_agent_state, **args) # 访问各报告: final_state["market_report"] final_state["sentiment_report"] final_state["news_report"] final_state["fundamentals_report"] ``` - **日志文件**: `tradingagents/graph/trading_graph.py` 第 195-225 行 - 保存位置: `eval_results/{ticker}/TradingAgentsStrategy_logs/full_states_log_{date}.json` - 包含所有报告字段 - **Agent实现**: 查看各分析师如何写入报告 - Market: `tradingagents/agents/analysts/market_analyst.py` 第 80-83 行 - Social: `tradingagents/agents/analysts/social_media_analyst.py` - News: `tradingagents/agents/analysts/news_analyst.py` - Fundamentals: `tradingagents/agents/analysts/fundamentals_analyst.py` --- ### **阶段 2: 研究辩论阶段 (Research Debate Phase)** #### 输入检查: - **从阶段1接收**: `final_state["market_report"]`, `final_state["sentiment_report"]`, etc. - **初始辩论状态**: `tradingagents/graph/propagation.py` 第 26-28 行 ```python "investment_debate_state": { "history": "", "current_response": "", "count": 0 } ``` #### 输出检查: - **运行时**: 在 `graph.stream()` 的 chunk 中检查 ```python if "investment_debate_state" in chunk: debate_state = chunk["investment_debate_state"] print(f"Bull history: {debate_state['bull_history']}") print(f"Bear history: {debate_state['bear_history']}") print(f"Judge decision: {debate_state['judge_decision']}") ``` - **最终状态**: ```python final_state["investment_debate_state"]["bull_history"] final_state["investment_debate_state"]["bear_history"] final_state["investment_debate_state"]["judge_decision"] final_state["investment_plan"] # Research Manager 的输出 final_state["trader_investment_plan"] # Trader 的输出 ``` - **日志文件**: `tradingagents/graph/trading_graph.py` 第 204-214 行 ```json { "investment_debate_state": { "bull_history": "...", "bear_history": "...", "history": "...", "current_response": "...", "judge_decision": "..." }, "trader_investment_decision": "..." } ``` - **Agent实现**: - Bull Researcher: `tradingagents/agents/researchers/bull_researcher.py` - Bear Researcher: `tradingagents/agents/researchers/bear_researcher.py` - Research Manager: `tradingagents/agents/managers/research_manager.py` - Trader: `tradingagents/agents/trader/trader.py` --- ### **阶段 3: 风险分析阶段 (Risk Analysis Phase)** #### 输入检查: - **从阶段2接收**: `final_state["trader_investment_plan"]` - **初始风险状态**: `tradingagents/graph/propagation.py` 第 29-37 行 ```python "risk_debate_state": { "history": "", "current_risky_response": "", "current_safe_response": "", "current_neutral_response": "", "count": 0 } ``` #### 输出检查: - **运行时**: 在 `graph.stream()` 的 chunk 中检查 ```python if "risk_debate_state" in chunk: risk_state = chunk["risk_debate_state"] print(f"Risky history: {risk_state['risky_history']}") print(f"Safe history: {risk_state['safe_history']}") print(f"Neutral history: {risk_state['neutral_history']}") print(f"Judge decision: {risk_state['judge_decision']}") ``` - **最终状态**: ```python final_state["risk_debate_state"]["risky_history"] final_state["risk_debate_state"]["safe_history"] final_state["risk_debate_state"]["neutral_history"] final_state["risk_debate_state"]["judge_decision"] final_state["final_trade_decision"] # 最终交易决策 ``` - **日志文件**: `tradingagents/graph/trading_graph.py` 第 216-222 行 ```json { "risk_debate_state": { "risky_history": "...", "safe_history": "...", "neutral_history": "...", "history": "...", "judge_decision": "..." }, "final_trade_decision": "..." } ``` - **Agent实现**: - Risky Analyst: `tradingagents/agents/risk_mgmt/aggresive_debator.py` - Safe Analyst: `tradingagents/agents/risk_mgmt/conservative_debator.py` - Neutral Analyst: `tradingagents/agents/risk_mgmt/neutral_debator.py` - Risk Manager: `tradingagents/agents/managers/risk_manager.py` --- ## 🛠️ 实际使用示例 ### 方法 1: 在代码中检查 (推荐用于调试) ```python from tradingagents.graph.trading_graph import TradingAgentsGraph from tradingagents.default_config import DEFAULT_CONFIG ta = TradingAgentsGraph(debug=True, config=DEFAULT_CONFIG.copy()) # 运行分析 final_state, decision = ta.propagate("NVDA", "2024-05-10") # 检查阶段1输出 print("=== 阶段1: 分析师报告 ===") print(f"Market Report: {final_state['market_report']}") print(f"Sentiment Report: {final_state['sentiment_report']}") print(f"News Report: {final_state['news_report']}") print(f"Fundamentals Report: {final_state['fundamentals_report']}") # 检查阶段2输出 print("\n=== 阶段2: 投资辩论 ===") debate = final_state['investment_debate_state'] print(f"Bull History: {debate['bull_history']}") print(f"Bear History: {debate['bear_history']}") print(f"Judge Decision: {debate['judge_decision']}") print(f"Investment Plan: {final_state['investment_plan']}") print(f"Trader Plan: {final_state['trader_investment_plan']}") # 检查阶段3输出 print("\n=== 阶段3: 风险分析 ===") risk = final_state['risk_debate_state'] print(f"Risky History: {risk['risky_history']}") print(f"Safe History: {risk['safe_history']}") print(f"Neutral History: {risk['neutral_history']}") print(f"Risk Judge Decision: {risk['judge_decision']}") print(f"Final Trade Decision: {final_state['final_trade_decision']}") ``` ### 方法 2: 查看日志文件 运行后，检查 JSON 日志文件: ```bash cat eval_results/NVDA/TradingAgentsStrategy_logs/full_states_log_2024-05-10.json ``` ### 方法 3: 使用 Debug 模式实时查看 在 `trading_graph.py` 第 171-179 行，debug 模式会打印每个节点的输出: ```python for chunk in self.graph.stream(init_agent_state, **args): if len(chunk["messages"]) == 0: pass else: chunk["messages"][-1].pretty_print() # 打印消息 # 可以在这里检查 chunk 中的其他字段 ``` ### 方法 4: 使用 CLI 界面 运行 CLI 可以看到实时输出: ```bash python -m cli.main ``` CLI 会显示每个阶段的进度和输出 (`cli/main.py` 第 888-923 行处理各阶段的输出显示) --- ## 📊 数据流总结 ``` 输入 → 阶段1 → 阶段2 → 阶段3 → 输出 ↓ ↓ ↓ reports debate risk final_decision ``` - **输入**: `propagation.py` 的 `create_initial_state()` - **阶段1输出**: `AgentState` 中的 `*_report` 字段 - **阶段2输出**: `AgentState` 中的 `investment_debate_state` 和 `trader_investment_plan` - **阶段3输出**: `AgentState` 中的 `risk_debate_state` 和 `final_trade_decision` - **最终日志**: `eval_results/{ticker}/TradingAgentsStrategy_logs/full_states_log_{date}.json`