Fix:修复API验证时推理模型大概率无法通过验证的问题

backend/app/core/agents/agent.py

@@ -1,4 +1,5 @@
 from app.core.llm.llm import LLM, simple_chat
+from app.config.setting import settings
 from app.utils.log_util import logger
 from icecream import ic
 
@@ -11,7 +12,7 @@ def __init__(
         self,
         task_id: str,
         model: LLM,
-        max_chat_turns: int = 30,  # 单个agent最大对话轮次
+        max_chat_turns: int = settings.MAX_CHAT_TURNS,  # 单个agent最大对话轮次
         max_memory: int = 12,  # 最大记忆轮次
     ) -> None:
         self.task_id = task_id

backend/app/core/agents/coder_agent.py

@@ -1,3 +1,4 @@
+import json_repair
 from app.core.agents.agent import Agent
 from app.config.setting import settings
 from app.utils.log_util import logger
@@ -8,7 +9,6 @@
 from app.schemas.A2A import CoderToWriter
 from app.core.prompts import CODER_PROMPT
 from app.utils.common_utils import get_current_files
-import json
 from app.core.prompts import get_reflection_prompt
 from app.core.functions import coder_tools
 
@@ -113,7 +113,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                             ),
                         )
 
-                        code = json.loads(tool_call.function.arguments)["code"]
+                        code = json_repair.loads(tool_call.function.arguments)["code"]
 
                         await redis_manager.publish_message(
                             self.task_id,
@@ -173,6 +173,7 @@ async def run(self, prompt: str, subtask_title: str) -> CoderToWriter:
                                     "content": text_to_gpt,
                                 }
                             )
+                            retry_count = 0
                             # 成功执行后继续循环，等待下一步指令
                             continue
                 else:

backend/app/core/agents/coordinator_agent.py

@@ -1,8 +1,10 @@
+import json
+import re
+from json_repair import repair_json
+from app.config.setting import settings
 from app.core.agents.agent import Agent
 from app.core.llm.llm import LLM
 from app.core.prompts import COORDINATOR_PROMPT
-import json
-import re
 from app.utils.log_util import logger
 from app.schemas.A2A import CoordinatorToModeler
 
@@ -12,7 +14,7 @@ def __init__(
         self,
         task_id: str,
         model: LLM,
-        max_chat_turns: int = 30,
+        max_chat_turns: int = settings.MAX_CHAT_TURNS,
     ) -> None:
         super().__init__(task_id, model, max_chat_turns)
         self.system_prompt = COORDINATOR_PROMPT
@@ -39,8 +41,8 @@ async def run(self, ques_all: str) -> CoordinatorToModeler:
 
                 if not json_str:
                     raise ValueError("返回的 JSON 字符串为空")
-
-                questions = json.loads(json_str)
+            
+                questions = json.loads(repair_json(json_str))
                 ques_count = questions["ques_count"]
                 logger.info(f"questions:{questions}")
                 return CoordinatorToModeler(questions=questions, ques_count=ques_count)

backend/app/core/agents/modeler_agent.py

@@ -1,10 +1,11 @@
 from app.core.agents.agent import Agent
 from app.core.llm.llm import LLM
+from app.config.setting import settings
 from app.core.prompts import MODELER_PROMPT
 from app.schemas.A2A import CoordinatorToModeler, ModelerToCoder
-from app.utils.log_util import logger
-import json
 from icecream import ic
+from json_repair import repair_json
+import json
 
 # TODO: 提问工具tool
 
@@ -14,7 +15,7 @@ def __init__(
         self,
         task_id: str,
         model: LLM,
-        max_chat_turns: int = 30,  # 添加最大对话轮次限制
+        max_chat_turns: int = settings.MAX_CHAT_TURNS,  # 添加最大对话轮次限制
     ) -> None:
         super().__init__(task_id, model, max_chat_turns)
         self.system_prompt = MODELER_PROMPT
@@ -42,7 +43,7 @@ async def run(self, coordinator_to_modeler: CoordinatorToModeler) -> ModelerToCo
         if not json_str:
             raise ValueError("返回的 JSON 字符串为空，请检查输入内容。")
         try:
-            questions_solution = json.loads(json_str)
+            questions_solution = json.loads(repair_json(json_str))
             ic(questions_solution)
             return ModelerToCoder(questions_solution=questions_solution)
         except json.JSONDecodeError as e:

backend/app/core/agents/writer_agent.py

@@ -1,12 +1,13 @@
+import json
 from app.core.agents.agent import Agent
 from app.core.llm.llm import LLM
+from app.config.setting import settings
 from app.core.prompts import get_writer_prompt
 from app.schemas.enums import CompTemplate, FormatOutPut
 from app.tools.openalex_scholar import OpenAlexScholar
 from app.utils.log_util import logger
 from app.services.redis_manager import redis_manager
 from app.schemas.response import SystemMessage, WriterMessage
-import json
 from app.core.functions import writer_tools
 from icecream import ic
 from app.schemas.A2A import WriterResponse
@@ -21,7 +22,7 @@ def __init__(
         self,
         task_id: str,
         model: LLM,
-        max_chat_turns: int = 10,  # 添加最大对话轮次限制
+        max_chat_turns: int = settings.MAX_CHAT_TURNS,  # 添加最大对话轮次限制
         comp_template: CompTemplate = CompTemplate,
         format_output: FormatOutPut = FormatOutPut.Markdown,
         scholar: OpenAlexScholar = None,

backend/app/core/prompts.py

@@ -94,7 +94,7 @@
 1. Primary: Seaborn (Nature/Science style)
 2. Secondary: Matplotlib
 3. Always:
-   - Handle Chinese characters properly
+   - Handle Chinese characters properly,for example,use SimHei font.
    - Set semantic filenames (e.g., "feature_correlation.png")
    - Save figures to working directory
    - Include model evaluation printouts
@@ -119,7 +119,8 @@
 - Release unused resources immediately
 
 
-Key improvements:
+Key improvements
+0. Do not generate or fabricate data. If there is an error where the file cannot be found, please search for the data from the question or the current working directory
 1. **Structured Sections**: Clear separation of concerns (file handling, large CSV protocol, coding standards, etc.)
 2. **Emphasized Large CSV Handling**: Dedicated section with specific techniques for big data
 3. **Optimized Readability**: Bulleted lists and code examples for quick scanning

backend/app/routers/modeling_router.py

@@ -98,7 +98,6 @@ async def validate_api_key(request: ValidateApiKeyRequest):
         await litellm.acompletion(
             model=request.model_id,
             messages=[{"role": "user", "content": "Hi"}],
-            max_tokens=1,
             api_key=request.api_key,
             base_url=request.base_url
             if request.base_url != "https://api.openai.com/v1"

backend/pyproject.toml

@@ -14,6 +14,7 @@ dependencies = [
     "httpx[socks]>=0.28.1",
     "icecream>=2.1.4",
     "ipykernel>=6.29.5",
+    "json-repair>=0.52.1",
     "jupyter-client>=8.6.3",
     "litellm>=1.69.0",
     "loguru>=0.7.3",