Merge pull request #3 from jihe520/dev

Dev

Author: jihe520

README.md

@@ -1 +1,105 @@
-## Test
+<h1 align="center">🤖 MathModelAgent 📐</h1>
+<p align="center">
+    <img src="./docs/icon.png" height="250px">
+</p>
+<h4 align="center">
+    专为数学建模设计的 Agent<br>
+    自动完成数学建模，生成一份完整的可以直接提交的论文。
+</h4>
+
+<h5 align="center">简体中文 | <a href="README_EN.md">English</a></h5>
+
+## 🌟 愿景：
+
+3 天的比赛时间变为 1 小时 <br> 
+自动完整一份可以获奖级别的建模论文
+
+<p align="center">
+    <img src="./docs/webui.png">
+</p>
+
+## ✨ 功能特性
+
+- 🔍 自动分析问题，数学建模，编写代码，纠正错误，撰写论文
+- 💻 本地代码解释器
+- 📝 生成一份编排好格式的论文
+- 🤝 muti-agents: ~~建模手~~，代码手(反思模块，本地代码解释器)，论文手
+- 🔄 muti-llms: 每个agent设置不同的模型
+- 💰 成本低 agentless(单次任务成本约 1 rmb)
+
+## 🚀 后期计划
+
+- [x] 添加并完成 webui、cli
+- [ ] 完善的教程、文档
+- [ ] 提供 web 服务
+- [ ] 英文支持（美赛）
+- [ ] 集成 latex 模板
+- [ ] 接入视觉模型
+- [ ] 添加正确文献引用
+- [ ] 更多测试案例
+- [ ] docker 部署
+- [ ] 引入用户的交互（选择模型，重写等等）
+- [ ] codeinterpreter 接入云端 如 e2b 等供应商..
+- [ ] 多语言: R 语言, matlab
+
+clone 项目后，下载**Todo Tree**插件，可以查看代码中所有具体位置的 todo
+
+## 视频demo
+
+<video src="https://github.com/user-attachments/assets/10b3145a-feb7-4894-aaca-30d44bb35b9e"></video>
+
+## 📖 使用教程
+
+1. 安装依赖
+
+```bash
+git clone https://github.com/jihe520/MathModelAgent.git # 克隆项目
+pip install uv # 推荐使用 uv 管理 python 项目
+uv venv # 创建虚拟环境
+uv sync # 安装依赖
+```
+
+2. 配置模型
+
+复制`/config/config.toml.example`到`/config/config.toml`, 填写配置模型和 APIKEY
+[deepseek开发者平台](https://platform.deepseek.com/)
+
+推荐模型能力较强的、参数量大的模型。
+
+3. 运行测试 和 启动项目
+
+```bash
+uv run example.py # 简单测试能否正确运行
+uv run webui.py # 推荐: 启动 webui
+uv run terminal.py # 启动终端项目
+```
+
+运行的结果在`/project/work_dir/`目录下
+
+## 🤝 贡献
+
+- 项目处于**开发阶段**（我有时间就会更新），变更较多，还存在许多 Bug，我正着手修复。
+- 欢迎提交 issues 和 PRs
+- 需求参考 后期计划
+
+## 📄 版权License
+
+请勿商业用途，商业用途联系我（作者）
+
+## 🙏 Reference
+
+Thanks to the following projects:
+- [OpenCodeInterpreter](https://github.com/OpenCodeInterpreter/OpenCodeInterpreter/tree/main)
+- [TaskWeaver](https://github.com/microsoft/TaskWeaver)
+- [Code-Interpreter](https://github.com/MrGreyfun/Local-Code-Interpreter/tree/main)
+- [Latex](https://github.com/Veni222987/MathModelingLatexTemplate/tree/main)
+
+
+## 其他
+
+感谢赞助
+[danmo-tyc](https://github.com/danmo-tyc)
+
+有问题可以进群问
+
+<img src="./docs/qq.jpg" height="400px">

README_EN.md

@@ -0,0 +1,82 @@
+<h1 align="center">🤖 MathModelAgent 📐</h1>
+<p align="center">
+    <img src="./docs/icon.png" height="250px">
+</p>
+<h4 align="center">
+    An agent designed for mathematical modeling<br>
+    Automatically complete mathematical modeling and generate a ready-to-submit paper.
+</h4>
+
+<br>
+<h5 align="center"> <a href="README.md">简体中文</a> | English </h5>
+
+## Vision:
+3 days to 1 hour <br>
+Automatically generate an award-worthy mathematical modeling paper
+
+## ✨ Features
+
+- 🔍 Automatic problem analysis, mathematical modeling, code writing, error correction, and paper writing
+- 💻 Local code interpreter
+- 📝 Generate a well-formatted paper
+- 🤝 Multi-agents: ~~modeling expert~~, coding expert (reflection module, local code interpreter), paper expert
+- 🔄 Multi-llms: Different models for each agent
+
+## 🚀 Future Plans
+
+- [ ] Comprehensive tutorials and documentation
+- [ ] Web service
+- [ ] English support (MCM/ICM)
+- [ ] LaTeX template integration
+- [ ] Vision model integration
+- [ ] Proper citation implementation
+- [ ] More test cases
+- [ ] Docker deployment
+- [ ] User interaction (model selection, rewriting, etc.)
+- [ ] Cloud integration for code interpreter (e.g., e2b providers)
+
+After cloning the project, install the **Todo Tree** plugin to view all specific todo locations in the code
+
+## Video Demo
+
+<video src="https://github.com/user-attachments/assets/10b3145a-feb7-4894-aaca-30d44bb35b9e"></video>
+
+
+## 📖 Usage Guide
+
+1. Install Dependencies
+```bash
+git clone https://github.com/jihe520/MathModelAgent.git
+pip install uv
+uv venv # create venv
+uv sync # install dependencies
+```
+2. Configure Model
+Copy `/config/config.toml.example` to `/config/config.toml` and fill in the model configuration
+
+Recommend using models with strong capabilities and large parameter counts.
+
+3. Run Tests and Start Project
+```bash
+uv run example.py # simple test to verify correct operation
+uv run main.py # complete project
+```
+
+Results will be in the `/project/work_dir/` directory
+
+## 🤝 Contribution
+
+- Project is in **development stage** (I update when time permits), with frequent changes and some bugs that I'm working to fix.
+- Issues and PRs are welcome
+- For requirements, refer to Future Plans
+
+## 📄 License
+
+Not for commercial use. Contact me (author) for commercial purposes.
+
+## 🙏 Reference
+Thanks to the following projects:
+- [OpenCodeInterpreter](https://github.com/OpenCodeInterpreter/OpenCodeInterpreter/tree/main)
+- [TaskWeaver](https://github.com/microsoft/TaskWeaver)
+- [Code-Interpreter](https://github.com/MrGreyfun/Local-Code-Interpreter/tree/main)
+- [Latex](https://github.com/Veni222987/MathModelingLatexTemplate/tree/main)

backend/app/core/agents.py

@@ -34,7 +34,7 @@ def __init__(
         self.current_chat_turns = 0  # 当前对话轮次计数器
         self.user_output = user_output
 
-    async def run(self, prompt: str, system_prompt: str) -> str:
+    async def run(self, prompt: str, system_prompt: str, sub_title: str) -> str:
         """
         执行agent的对话并返回结果和总结
 
@@ -54,7 +54,9 @@ async def run(self, prompt: str, system_prompt: str) -> str:
 
             # 获取历史消息用于本次对话
             response = await self.model.chat(
-                history=self.chat_history, agent_name=self.__class__.__name__
+                history=self.chat_history,
+                agent_name=self.__class__.__name__,
+                sub_title=sub_title,
             )
             response_content = response.choices[0].message.content
             self.chat_history.append({"role": "assistant", "content": response_content})
@@ -140,7 +142,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                 f"Failed to complete task after {self.max_retries} attempts. Last error: {last_error_message}"
             )
 
-            # try:
+        # try:
         while (
             not task_completed
             and retry_count < self.max_retries
@@ -189,6 +191,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                     )
 
                     # 执行工具调用
+                    logger.warning("执行工具调用")
                     (
                         text_to_gpt,
                         error_occurred,
@@ -207,6 +210,8 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                     )
 
                     if error_occurred:
+                        logger.warning("代码执行错误")
+
                         retry_count += 1
                         last_error_message = error_message
                         reflection_prompt = get_reflection_prompt(error_message, code)
@@ -224,6 +229,8 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                     # 检查任务完成情况时也计入对话轮次
                     self.current_chat_turns += 1
                     # 使用所有执行结果生成检查提示
+                    logger.warning("判断是否完成")
+
                     completion_check_prompt = get_completion_check_prompt(
                         prompt, text_to_gpt
                     )
@@ -245,6 +252,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                         hasattr(completion_response.choices[0].message, "tool_calls")
                         and completion_response.choices[0].message.tool_calls
                     ):
+                        logger.warning("没有调用工具，代表已经完成了")
                         task_completed = True
                         return completion_response.choices[0].message.content
 
@@ -282,28 +290,25 @@ async def run(
         self,
         prompt: str,
         available_images: list[str] = None,
-        static_prefix: str = "/static/",
+        sub_title: str = None,
     ) -> str:
         """
         执行写作任务
         Args:
             prompt: 写作提示
             available_images: 可用的图片相对路径列表（如 20250420-173744-9f87792c/编号_分布.png）
-            static_prefix: 静态资源前缀
+            sub_title: 子任务标题
         """
+        logger.info(f"subtitle是:{sub_title}")
+
         if available_images:
             self.available_images = available_images
             # 拼接成完整URL
-            image_list = "\n".join(
-                [
-                    f"- {static_prefix}{img if img.startswith('/') else '/' + img}"
-                    for img in available_images
-                ]
-            )
+            image_list = ",".join(available_images)
             image_prompt = f"\n可用的图片链接列表：\n{image_list}\n请在写作时适当引用这些图片链接。"
             prompt = prompt + image_prompt
 
-        return await super().run(prompt, self.system_prompt)
+        return await super().run(prompt, self.system_prompt, sub_title)
 
     async def summarize(self) -> str:
         """

backend/app/core/llm.py

@@ -1,10 +1,12 @@
 import json
+from app.utils.common_utils import transform_link
 from openai import OpenAI
 from app.utils.log_util import logger
 import time
 from app.schemas.response import AgentMessage, CoderMessage, WriterMessage
 from app.utils.enums import AgentType
 from app.utils.redis_manager import redis_manager
+import re
 
 
 class LLM:
@@ -32,7 +34,10 @@ async def chat(
         retry_delay: float = 1.0,  # 添加重试延迟
         top_p: float | None = None,  # 添加top_p参数,
         agent_name: str = "NO_NAME",  # CoderAgent or WriterAgent
+        sub_title: str | None = None,
     ) -> str:
+        logger.info(f"subtitle是:{sub_title}")
+
         kwargs = {
             "model": self.model,
             "messages": history,
@@ -51,10 +56,11 @@ async def chat(
         for attempt in range(max_retries):
             try:
                 completion = self.client.chat.completions.create(**kwargs)
+                logger.info(f"API返回: {completion}")
                 if not completion or not hasattr(completion, "choices"):
                     raise ValueError("无效的API响应")
                 self.chat_count += 1
-                await self.analyse_completion(completion, agent_name)
+                await self.analyse_completion(completion, agent_name, sub_title)
                 return completion
             except json.JSONDecodeError:
                 logger.error(f"第{attempt + 1}次重试: API返回无效JSON")
@@ -64,7 +70,9 @@ async def chat(
                 logger.debug(f"请求参数: {kwargs}")
                 raise  # 如果所有重试都失败，则抛出异常
 
-    async def analyse_completion(self, completion, agent_name):
+    async def analyse_completion(self, completion, agent_name, sub_title):
+        logger.info(f"subtitle是:{sub_title}")
+
         code = ""
         if (
             hasattr(completion.choices[0].message, "tool_calls")
@@ -73,13 +81,22 @@ async def analyse_completion(self, completion, agent_name):
             tool_call = completion.choices[0].message.tool_calls[0]
             if tool_call.function.name == "execute_code":
                 code = json.loads(tool_call.function.arguments)["code"]
-        await self.send_message(agent_name, completion.choices[0].message.content, code)
+        (
+            await self.send_message(
+                agent_name, completion.choices[0].message.content, code, sub_title
+            )
+        )
 
-    async def send_message(self, agent_name, content, code=""):
+    async def send_message(self, agent_name, content, code="", sub_title=None):
+        logger.info(f"subtitle是:{sub_title}")
         if agent_name == "CoderAgent":
             agent_msg: CoderMessage = CoderMessage(content=content, code=code)
         elif agent_name == "WriterAgent":
-            agent_msg: WriterMessage = WriterMessage(content=content)
+            # 处理 Markdown 格式的图片语法
+            content = transform_link(self.task_id, content)
+            agent_msg: WriterMessage = WriterMessage(
+                content=content, sub_title=sub_title
+            )
         else:
             raise ValueError(f"无效的agent_name: {agent_name}")
 

backend/app/core/prompts.py

@@ -75,12 +75,12 @@ def get_writer_prompt(
         skill：熟练掌握{format_output}排版,
         output：你需要按照要求的格式排版,只输出{format_output}排版的内容
         
-        1. 当你输入图像引用时候，你需要将用户输入的文件名称路径切换为相对路径
-        如用户输入文件路径image_name.png,你转化为../jupyter/image_name.png,就可正确引用显示
-        2. 你不需要输出markdown的这个```markdown格式，只需要输出markdown的内容
-        3. 严格按照参考用户输入的格式模板以及**正确的编号顺序**
-        4. 不需要询问用户
-        5. 当提到图片时，请使用提供的图片列表中的文件名
+        1. 当你输入图像引用时候，使用![image_name](image_name.png)
+        2. 你不需要输出markdown的这个```markdown格式，只需要输出markdown的内容，
+        3. Latex公式使用$$ $$包裹 
+        4. 严格按照参考用户输入的格式模板以及**正确的编号顺序**
+        5. 不需要询问用户 
+        6. 当提到图片时，请使用提供的图片列表中的文件名
         """
 
 

backend/app/core/workflow.py

@@ -126,6 +126,7 @@ async def execute(self, problem: Problem):
             writer_response = await writer_agent.run(
                 writer_prompt,
                 available_images=await e2b_code_interpreter.get_created_images(key),
+                sub_title=key,
             )
 
             await redis_manager.publish_message(
@@ -156,7 +157,7 @@ async def execute(self, problem: Problem):
                 comp_template=problem.comp_template,
                 format_output=problem.format_output,
             )
-            writer_response = await writer_agent.run(value)
+            writer_response = await writer_agent.run(prompt=value, sub_title=key)
             user_output.set_res(key, writer_response)
 
         logger.info(user_output.get_res())