一个基于 AI Agent 的逻辑审计与事实核查工具，通过拆解论证逻辑、联网搜索验证核心事实并进行多级信源评估，最终给出论证的可靠性评分。
-
配置环境: 复制
.env.example为.env并填写必要的 API Key (OpenAI, Tavily)。cp .env.example .env
-
安装依赖:
pip install -r requirements.txt
-
运行程序:
python main.py <待分析的文件路径>
graph TD
%% --- 样式定义 ---
classDef terminal fill:#2d3436,stroke:#2d3436,color:#fff,rx:10,ry:10;
classDef process fill:#00b894,stroke:#00b894,color:#fff,rx:5,ry:5;
classDef userAction fill:#6c5ce7,stroke:#6c5ce7,color:#fff,rx:5,ry:5;
classDef loopAction fill:#ff7675,stroke:#d63031,color:#fff,stroke-width:2px,stroke-dasharray: 5 5,rx:5,ry:5;
classDef decision fill:#ffeaa7,stroke:#fdcb6e,color:#2d3436,rx:5,ry:5;
%% --- 主流程 ---
Start([开始]):::terminal --> Decompose[1. 逻辑拆解 & 提取事实]:::process
Decompose --> CheckFeedback{用户反馈?}:::decision
CheckFeedback -- 修正 --> Decompose
CheckFeedback -- 确认 --> SelectFacts[2. 用户选择事实]:::userAction
SelectFacts --> PrepareFact((开始核查)):::terminal
%% --- 核心循环区域 ---
subgraph VerifyLoop ["3. 智能核查闭环 (Agentic Loop)"]
direction TB
PrepareFact --> Search[Tavily 联网搜索]:::loopAction
Search --> Reflect[LLM: 信源分级 & 评分]:::loopAction
Reflect --> IsSufficient{证据充足?}:::decision
IsSufficient -- "No (重试 < 3次)" --> Refine[生成更精准关键词]:::loopAction
Refine --> Search
IsSufficient -- "Yes / 次数耗尽" --> GenReport[生成单条核查报告]:::process
end
%% --- 结尾 ---
GenReport --> FinalEval[4. 最终可靠性评估]:::process
FinalEval --> End([结束]):::terminal