add search paper tool

add search paper tool

Author: jihe520

README.md

@@ -72,7 +72,7 @@
 ```bash
 # support all model, check out https://docs.litellm.ai/docs/ 
 API_KEY=
-# gpt-4.1,deepseek/deepseek-chat
+# gpt-4.1,deepseek/deepseek-chat,gemini/gemini-2.5-flash-preview-04-17
 MODEL=
 # 确保安装 Redis
 ```

backend/.env.dev.example

@@ -3,6 +3,7 @@ ENV=dev
 API_KEY=
 # gpt-4.1,deepseek/deepseek-chat
 MODEL=
+# BASE_URL= 不需要填
 
 # 模型最大问答次数
 MAX_CHAT_TURNS=60

backend/app/config/md_template.toml

@@ -29,8 +29,8 @@ firstPage = """
 RepeatQues = """
 # 一、问题重述
 ## 1.1 问题背景
-
-根据题目总结问题背景
+查询文献
+根据题目总结问题背景（查询文献）
 格式：一段话
 大约160字
 ## 1.2 问题重述

backend/app/config/setting.py

@@ -29,6 +29,7 @@ class Settings(BaseSettings):
     REDIS_MAX_CONNECTIONS: int
     CORS_ALLOW_ORIGINS: Annotated[list[str] | str, BeforeValidator(parse_cors)]
     SERVER_HOST: str = "http://localhost:8000"  # 默认值
+    OPENALEX_EMAIL: Optional[str] = None
 
     model_config = SettingsConfigDict(
         env_file=".env.dev",

backend/app/core/agents.py

@@ -18,6 +18,7 @@
 from app.schemas.response import SystemMessage
 from app.tools.base_interpreter import BaseCodeInterpreter
 from app.tools.openalex_scholar import OpenAlexScholar
+from icecream import ic
 
 
 class Agent:
@@ -78,9 +79,9 @@ def append_chat_history(self, msg: dict) -> None:
         # )
 
     def clear_memory(self):
-        logger.info(f"{self.__class__.__name__}:清除记忆")
         if len(self.chat_history) <= self.max_memory:
             return
+        logger.info(f"{self.__class__.__name__}:清除记忆")
 
         # 使用切片保留第一条和最后两条消息
         self.chat_history = self.chat_history[:2] + self.chat_history[-5:]
@@ -348,6 +349,7 @@ async def run(
             # 拼接成完整URL
             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。"
+            ic(image_prompt)
             prompt = prompt + image_prompt
 
         logger.info(f"{self.__class__.__name__}:开始:执行对话")

backend/app/core/llm.py

@@ -1,13 +1,11 @@
 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.schemas.response import CoderMessage, WriterMessage
 from app.utils.redis_manager import redis_manager
-import re
 from litellm import acompletion
+import litellm
 
 
 class LLM:
@@ -67,8 +65,8 @@ async def chat(
                 self.chat_count += 1
                 await self.analyse_completion(response, agent_name, sub_title)
                 return response
-            except json.JSONDecodeError:
-                logger.error(f"第{attempt + 1}次重试: API返回无效JSON")
+            except (json.JSONDecodeError, litellm.InternalServerError) as e:
+                logger.error(f"第{attempt + 1}次重试: {str(e)}")
                 if attempt < max_retries - 1:  # 如果不是最后一次尝试
                     time.sleep(retry_delay * (attempt + 1))  # 指数退避
                     continue
@@ -86,11 +84,10 @@ async def analyse_completion(self, completion, agent_name, sub_title):
             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, sub_title
             )
-        )
+        # 处理文献查询
 
     async def send_message(self, agent_name, content, code="", sub_title=None):
         logger.info(f"subtitle是:{sub_title}")

backend/app/core/prompts.py

@@ -15,7 +15,7 @@
 
 CODER_PROMPT = """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,numpy,pandas,matplotlib,seaborn,scikit-learn,xgboost,scipy and how to use their models, classes and functions.you can use them to do mathmodel and data analysis.
+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.
 
 
 When generating code:
@@ -25,7 +25,7 @@
 4. The working directory is already set up, and any uploaded files are already in the current directory
 5. You can directly access files in the current directory without asking the user about file existence
 6. For data analysis tasks, if you see Excel files (.xlsx), use pandas to read them directly
-7. try to visualize the data , process and  results using seaborn first and then matplotlibs,be nature sci style.
+7. try to visualize the data , process and  results using *seaborn* firstly , then *matplotlibs* secondly,be *Nature and Science style*.
 
 For example:
 # Correct:
@@ -85,11 +85,12 @@ def get_writer_prompt(
         
         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: 行内公式（Inline Formula） 和 块级公式（Block Formula
+        3. LaTex: 行内公式（Inline Formula） 和 块级公式（Block Formula）
         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
         7. when you write,check if you need to use tools search_papers to cite.if you need, markdown Footnote e.g.[^1]
+        8. 对于问题背景和模型介绍，需查询文献调用tools search_papers
         """
 
 

backend/app/core/workflow.py

@@ -34,7 +34,6 @@ async def execute(self, problem: Problem):
         self.task_id = problem.task_id
         self.work_dir = create_work_dir(self.task_id)
 
-        # default choose deepseek model
         llm_model = LLM(
             api_key=settings.API_KEY,
             model=settings.MODEL,
@@ -67,7 +66,6 @@ async def execute(self, problem: Problem):
         )
 
         # Example usage
-
         scholar = OpenAlexScholar(email=settings.OPENALEX_EMAIL)  # 请替换为您的真实邮箱
 
         await redis_manager.publish_message(
@@ -156,13 +154,6 @@ async def execute(self, problem: Problem):
                 SystemMessage(content=f"论文手开始写{key}部分"),
             )
 
-            # TODO: writer_agent 是否不需要初始化
-            writer_agent = WriterAgent(
-                task_id=problem.task_id,
-                model=llm_model,
-                comp_template=problem.comp_template,
-                format_output=problem.format_output,
-            )
             writer_response = await writer_agent.run(prompt=value, sub_title=key)
             user_output.set_res(key, writer_response)
 

backend/app/main.py

@@ -6,17 +6,18 @@
 from app.utils.log_util import logger
 from app.config.setting import settings
 from fastapi.staticfiles import StaticFiles
+from app.utils.cli import get_ascii_banner, center_cli_str
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
+    print(get_ascii_banner())
+    print(center_cli_str("GitHub:https://github.com/jihe520/MathModelAgent"))
     logger.info("Starting MathModelAgent")
 
     PROJECT_FOLDER = "./project"
     os.makedirs(PROJECT_FOLDER, exist_ok=True)
 
-    logger.info("CORS_ALLOW_ORIGINS:", settings.CORS_ALLOW_ORIGINS)
-
     yield
     logger.info("Stopping MathModelAgent")