Merge pull request #12 from jihe520/dev

Dev

Author: jihe520

backend/.env.dev.example

@@ -4,7 +4,7 @@ ENV=dev
 # 为每个 agent 配置合适的模型
 COORDINATOR_API_KEY=
 COORDINATOR_MODEL=
-# COORDINATOR_BASE_URL= 如果需要中转自定义等 
+# COORDINATOR_BASE_URL= 如果需要中转自定义等 地址后面添加 /v1
 
 # 推荐 thinking model
 MODELER_API_KEY=

backend/app/core/agents/coder_agent.py

@@ -11,6 +11,7 @@
 import json
 from app.core.prompts import get_reflection_prompt, get_completion_check_prompt
 from app.core.functions import coder_tools
+from icecream import ic
 
 
 # 代码强
@@ -54,7 +55,6 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
 
         retry_count = 0
         last_error_message = ""
-        task_completed = False
 
         if self.current_chat_turns >= self.max_chat_turns:
             logger.error(f"超过最大聊天次数: {self.max_chat_turns}")
@@ -78,8 +78,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
 
         # try:
         while (
-            not task_completed
-            and retry_count < self.max_retries
+            retry_count < self.max_retries
             and self.current_chat_turns < self.max_chat_turns
         ):
             self.current_chat_turns += 1
@@ -120,6 +119,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
 
                     # 更新对话历史 - 添加助手的响应
                     self.append_chat_history(response.choices[0].message.model_dump())
+                    logger.info(response.choices[0].message.model_dump())
 
                     # 执行工具调用
                     logger.info("执行工具调用")
@@ -167,47 +167,15 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                                 "content": text_to_gpt,
                             }
                         )
-
-                    # 检查任务完成情况时也计入对话轮次
-                    self.current_chat_turns += 1
-                    logger.info(
-                        f"当前对话轮次: {self.current_chat_turns} / {self.max_chat_turns}"
-                    )
-                    # 使用所有执行结果生成检查提示
-                    logger.info("判断是否完成")
-
-                    completion_check_prompt = get_completion_check_prompt(
-                        prompt, text_to_gpt
-                    )
-                    self.append_chat_history(
-                        {"role": "user", "content": completion_check_prompt}
-                    )
-
-                    completion_response = await self.model.chat(
-                        history=self.chat_history,
-                        tools=coder_tools,
-                        tool_choice="auto",
-                        agent_name=self.__class__.__name__,
-                    )
-
-                    ## 没有调用工具，代表已经完成了
-                    if not (
-                        hasattr(completion_response.choices[0].message, "tool_calls")
-                        and completion_response.choices[0].message.tool_calls
-                    ):
-                        logger.info("没有调用工具，代表任务已完成")
-                        task_completed = True
-                        return CoderToWriter(
-                            coder_response=completion_response.choices[
-                                0
-                            ].message.content,
-                            created_images=await self.code_interpreter.get_created_images(
-                                subtask_title
-                            ),
-                        )
             else:
-                logger.info("没有工具，代表任务完成")
-                task_completed = True
+                # 没有工具调用，表示任务完成
+                logger.info("没有工具调用，任务完成")
+                return CoderToWriter(
+                    coder_response=response.choices[0].message.content,
+                    created_images=await self.code_interpreter.get_created_images(
+                        subtask_title
+                    ),
+                )
 
             if retry_count >= self.max_retries:
                 logger.error(f"超过最大尝试次数: {self.max_retries}")

backend/app/core/agents/writer_agent.py

@@ -87,7 +87,6 @@ async def run(
             logger.info("检测到工具调用")
             tool_call = response.choices[0].message.tool_calls[0]
             tool_id = tool_call.id
-            tool_call.function.name
             if tool_call.function.name == "search_papers":
                 logger.info("调用工具: search_papers")
                 await redis_manager.publish_message(
@@ -104,24 +103,9 @@ async def run(
                     ),
                 )
 
-                full_content = response.choices[0].message.content
                 # 更新对话历史 - 添加助手的响应
-                self.append_chat_history(
-                    {
-                        "role": "assistant",
-                        "content": full_content,
-                        "tool_calls": [
-                            {
-                                "id": tool_id,
-                                "type": "function",
-                                "function": {
-                                    "name": "search_papers",
-                                    "arguments": json.dumps({"query": query}),
-                                },
-                            }
-                        ],
-                    }
-                )
+                self.append_chat_history(response.choices[0].message.model_dump())
+                ic(response.choices[0].message.model_dump())
 
                 try:
                     papers = await self.scholar.search_papers(query)

backend/app/core/prompts.py

@@ -1,9 +1,10 @@
 from app.schemas.enums import FormatOutPut
-
+import platform
 
 FORMAT_QUESTIONS_PROMPT = """
 用户将提供给你一段题目信息，**请你不要更改题目信息，完整将用户输入的内容**，以 JSON 的形式输出，输出的 JSON 需遵守以下的格式：
 
+```json
 {
   "title": <题目标题>      
   "background": <题目背景，用户输入的一切不在title，ques1，ques2，ques3...中的内容都视为问题背景信息background>,
@@ -12,6 +13,7 @@
   "ques2": <问题2>,
   "ques3": <问题3,用户输入的存在多少问题，就输出多少问题ques1,ques2,ques3...以此类推>,
 }
+```
 """
 
 
@@ -32,26 +34,35 @@
 skill：熟练掌握各种数学建模的模型和思路
 output：数学建模的思路和使用到的模型
 attention：不需要给出代码，只需要给出思路和模型
-format：以 JSON 的形式输出输出的 JSON,需遵守以下的格式：
+
+# 输出规范
+## 字段约束
+
+以 JSON 的形式输出输出的 JSON,需遵守以下的格式：
+```json
 {
   "eda": <数据分析EDA方案>,
   "ques1": <问题1的建模思路和模型方案>,
-  "ques2": <问题2的建模思路和模型方案>,
-  "ques3": <问题3的建模思路和模型方案,用户输入的存在多少问题，就输出多少问题ques1,ques2,ques3...以此类推>,
+  "quesN": <问题N的建模思路和模型方案>,
   "sensitivity_analysis": <敏感性分析方案>,
 }
-只需要以上 eda,ques1,ques2,ques3,ques.. ,sensitivity_analysis 方面建模思路，不需要要其他json key
-只需要 key value 的 dict，不要嵌套
-如果没有 ques num ，则不需要该 key
-用户可能会提出意见，你需要根据意见后，按格式修改建模思路
+```
+* 根据实际问题数量动态生成ques1,ques2...quesN
+
+## 输出约束
+- json key 只能是上面的: eda,ques1,quesN,sensitivity_analysis
+- 严格保持单层JSON结构
+- 键值对值类型：字符串
+- 禁止嵌套/多级JSON
 """
 
 # TODO : 对于特大 csv 读取
 
-CODER_PROMPT = """You are an AI code interpreter.
+CODER_PROMPT = f"""You are an AI code interpreter.
 Your goal is to help users do a variety of jobs by executing Python code.
 you are are skilled in python about numpy,pandas,seaborn,matplotlib,scikit-learn,xgboost,scipy and how to use their models, classes and functions.you can use them to do mathmodel and data analysis.
 
+environment:{platform.system()}
 
 When generating code:
 1. Use double quotes for strings containing Chinese characters
@@ -113,22 +124,43 @@ def get_writer_prompt(
     format_output: FormatOutPut = FormatOutPut.Markdown,
 ):
     return f"""
-        role：你是一名数学建模经验丰富的写作手，负责写作部分。
-        task: 根据问题和如下的模板写出解答,
-        skill：熟练掌握{format_output}排版,如图片、**公式**、表格、列表等
-        output：你需要按照要求的格式排版,只输出正确的{format_output}排版的内容
+        # Role Definition
+        Professional writer for mathematical modeling competitions with expertise in technical documentation and literature synthesis
+        
+        # Core Tasks
+        1. Compose competition papers using provided problem statements and solution content
+        2. Strictly adhere to {format_output} formatting templates
+        3. Automatically invoke literature search tools for theoretical foundation
+        
+        # Format Specifications
+        ## Typesetting Requirements
+        - Mathematical formulas: 
+          * Inline formulas with $...$ 
+          * Block formulas with $$...$$
+        - Visual elements: 
+          * Image references on new lines: ![alt_text](filename.ext)
+          * Table formatting with markdown syntax
+        - Citation system: 
+          * Direct inline citations with full bibliographic details
+          * Prohibit end-of-document reference lists
+
+        ## Citation Protocol
+        1. Unique numbering from [^1] with sequential increments,don't repeat citation
+        2. Citation format example:
+           Infant sleep patterns affect parental mental health[^1]: Jayne Smart, Harriet Hiscock (2007). Early infant crying and sleeping problems...
+        3. Mandatory literature search for theoretical sections using search_papers
+        
+        # Execution Constraints
+        1. Autonomous operation without procedural inquiries
+        2. Output pure {format_output} content without codeblock markers
+        3. Strict filename adherence for image references
+        4. Language consistency with user input (currently English)
         
-        1. When referencing an image, use ![image_name](image_name.png), and the image reference should be on a new line after the paragraph.
-        2. Do not output the ```markdown format; only output the markdown content itself.
-        3. For LaTeX: use $ for inline formulas and $$ for block formulas.
-        4. Strictly follow the reference user's format template and use the correct numbering order.
-        5. Don't ask the user anything about how to do or what to do next, just do it yourself.
-        6. When mentioning images, use the provided filenames from the image list.
-        7. Use markdown footnotes in related sentence, e.g. [^1].
-        8. List all used references at the end in markdown footnote format. Do not use a title #, just list them at the end.
-        9. Include an empty line between each citation for better readability.
-        10. For background and model introduction, you need to search the literature by calling tools search_papers.
-        11. Respond in the same language as the user.
+        # Exception Handling
+        Automatic tool invocation triggers:
+        1. Theoretical sections requiring references → search_papers
+        2. Methodology requiring diagrams → generate & insert after creation
+        3. Data interpretation needs → request analysis tools
         """