refactor(chat): 优化异步调用及增强内容审查机制

- 将chat_router中的predict_async重命名为call_async并调用model.call替代model.predict
- 在流式消息处理中新增加基于LLM的内容审查
- 配置类中增加enable_content_guard_llm及对应LLM模型配置项
- 兼容models.private.yaml替代旧的models.private.yml文件名
- chat_model及embedding模块统一将predict方法重命名为call或encode，增强接口语义
- ContentGuard新增基于LLM的内容合规检测功能，支持动态加载审查模型
- 更新静态模型配置提示，建议使用models.private.yaml文件
- 新增示例CSV测试数据文件，补充测试用例基础数据

server/routers/chat_router.py

@@ -86,11 +86,11 @@ async def call(query: str = Body(...), meta: dict = Body(None), current_user: Us
     meta = meta or {}
     model = select_model(model_provider=meta.get("model_provider"), model_name=meta.get("model_name"))
 
-    async def predict_async(query):
+    async def call_async(query):
         loop = asyncio.get_event_loop()
-        return await loop.run_in_executor(executor, model.predict, query)
+        return await loop.run_in_executor(executor, model.call, query)
 
-    response = await predict_async(query)
+    response = await call_async(query)
     logger.debug({"query": query, "response": response.content})
 
     return {"response": response.content}
@@ -169,15 +169,24 @@ async def chat_agent(
             async for msg, metadata in agent.stream_messages(messages, input_context=input_context):
                 # logger.debug(f"msg: {msg.model_dump()}, metadata: {metadata}")
                 if isinstance(msg, AIMessageChunk):
+
                     accumulated_content += msg.content
                     if conf.enable_content_guard and content_guard.check(accumulated_content):
                         logger.warning(f"Sensitive content detected in stream: {accumulated_content}")
                         yield make_chunk(message="检测到敏感内容，已中断输出", status="error")
                         return
+
                     yield make_chunk(content=msg.content, msg=msg.model_dump(), metadata=metadata, status="loading")
+
                 else:
                     yield make_chunk(msg=msg.model_dump(), metadata=metadata, status="loading")
 
+            # Additional content guard with llm
+            if conf.enable_content_guard and content_guard.check_with_llm(accumulated_content):
+                logger.warning(f"Content guard triggered: {accumulated_content=}")
+                yield make_chunk(message="检测到敏感内容，已中断输出", status="error")
+                return
+
             yield make_chunk(status="finished", meta=meta)
         except Exception as e:
             logger.error(f"Error streaming messages: {e}, {traceback.format_exc()}")

src/config/app.py

@@ -50,6 +50,12 @@ class Config(SimpleConfig):
         # 功能选项
         self.add_item("enable_reranker", default=False, des="是否开启重排序")
         self.add_item("enable_content_guard", default=False, des="是否启用内容审查")
+        self.add_item("enable_content_guard_llm", default=False, des="是否启用LLM内容审查")
+        self.add_item(
+            "content_guard_llm_model",
+            default="siliconflow/Qwen/Qwen3-235B-A22B-Instruct-2507",
+            des="内容审查LLM模型"
+        )
         self.add_item(
             "enable_web_search",
             default=False,
@@ -98,15 +104,21 @@ class Config(SimpleConfig):
 
     def _update_models_from_file(self):
         """
-        从 models.yaml 和 models.private.yml 中更新 MODEL_NAMES
+        从 models.yaml 和 models.private.yaml 中更新 MODEL_NAMES
         """
 
         with open(Path("src/static/models.yaml"), encoding="utf-8") as f:
             _models = yaml.safe_load(f)
 
-        # 尝试打开一个 models.private.yml 文件，用来覆盖 models.yaml 中的配置
+        # 尝试打开一个 models.private.yaml 文件，用来覆盖 models.yaml 中的配置
+        # 兼容性测试（历史遗留问题）
+        exist_yml = Path("src/static/models.private.yml").exists()
+        exist_yaml = Path("src/static/models.private.yaml").exists()
+        if exist_yml and not exist_yaml:
+            os.rename("src/static/models.private.yml", "src/static/models.private.yaml")
+
         try:
-            with open(Path("src/static/models.private.yml"), encoding="utf-8") as f:
+            with open(Path("src/static/models.private.yaml"), encoding="utf-8") as f:
                 _models_private = yaml.safe_load(f)
         except FileNotFoundError:
             _models_private = {}
@@ -124,7 +136,7 @@ class Config(SimpleConfig):
             "EMBED_MODEL_INFO": self.embed_model_names,
             "RERANKER_LIST": self.reranker_names,
         }
-        with open(Path("src/static/models.private.yml"), "w", encoding="utf-8") as f:
+        with open(Path("src/static/models.private.yaml"), "w", encoding="utf-8") as f:
             yaml.dump(_models, f, indent=2, allow_unicode=True)
 
     def handle_self(self):

src/models/chat_model.py

@@ -15,7 +15,7 @@ class OpenAIBase:
         self.model_name = model_name
         self.info = kwargs
 
-    def predict(self, message, stream=False):
+    def call(self, message, stream=False):
         if isinstance(message, str):
             messages = [{"role": "user", "content": message}]
         else:

src/models/embedding.py

@@ -29,30 +29,22 @@ class BaseEmbeddingModel(ABC):
         self.embed_state = {}
 
     @abstractmethod
-    def predict(self, message: list[str] | str) -> list[list[float]]:
+    def encode(self, message: list[str] | str) -> list[list[float]]:
         """同步编码"""
         raise NotImplementedError("Subclasses must implement this method")
 
+    def encode_queries(self, queries: list[str] | str) -> list[list[float]]:
+        """等同于encode"""
+        return self.encode(queries)
+
     @abstractmethod
-    async def apredict(self, message: list[str] | str) -> list[list[float]]:
+    async def aencode(self, message: list[str] | str) -> list[list[float]]:
         """异步编码"""
         raise NotImplementedError("Subclasses must implement this method")
 
-    def encode(self, message: list[str] | str) -> list[list[float]]:
-        """等同于predict"""
-        return self.predict(message)
-
-    def encode_queries(self, queries: list[str] | str) -> list[list[float]]:
-        """等同于predict"""
-        return self.predict(queries)
-
-    async def aencode(self, message: list[str] | str) -> list[list[float]]:
-        """等同于apredict"""
-        return await self.apredict(message)
-
     async def aencode_queries(self, queries: list[str] | str) -> list[list[float]]:
-        """等同于apredict"""
-        return await self.apredict(queries)
+        """等同于aencode"""
+        return await self.aencode(queries)
 
     def batch_encode(self, messages: list[str], batch_size: int = 40) -> list[list[float]]:
         # logger.info(f"Batch encoding {len(messages)} messages")
@@ -107,7 +99,7 @@ class OllamaEmbedding(BaseEmbeddingModel):
         super().__init__(**kwargs)
         self.base_url = self.base_url or get_docker_safe_url("http://localhost:11434/api/embed")
 
-    def predict(self, message: list[str] | str) -> list[list[float]]:
+    def encode(self, message: list[str] | str) -> list[list[float]]:
         if isinstance(message, str):
             message = [message]
 
@@ -123,7 +115,7 @@ class OllamaEmbedding(BaseEmbeddingModel):
             logger.error(f"Ollama Embedding request failed: {e}, {payload}")
             raise ValueError(f"Ollama Embedding request failed: {e}")
 
-    async def apredict(self, message: list[str] | str) -> list[list[float]]:
+    async def aencode(self, message: list[str] | str) -> list[list[float]]:
         if isinstance(message, str):
             message = [message]
 
@@ -149,7 +141,7 @@ class OtherEmbedding(BaseEmbeddingModel):
     def build_payload(self, message: list[str] | str) -> dict:
         return {"model": self.model, "input": message}
 
-    def predict(self, message: list[str] | str) -> list[list[float]]:
+    def encode(self, message: list[str] | str) -> list[list[float]]:
         payload = self.build_payload(message)
         try:
             response = requests.post(self.base_url, json=payload, headers=self.headers, timeout=60)
@@ -162,7 +154,7 @@ class OtherEmbedding(BaseEmbeddingModel):
             logger.error(f"Other Embedding request failed: {e}, {payload}")
             raise ValueError(f"Other Embedding request failed: {e}")
 
-    async def apredict(self, message: list[str] | str) -> list[list[float]]:
+    async def aencode(self, message: list[str] | str) -> list[list[float]]:
         payload = self.build_payload(message)
         async with httpx.AsyncClient() as client:
             try:

src/plugins/guard.py

@@ -1,5 +1,31 @@
 import os
 
+from src.config.app import config
+from src.models import select_model
+from src.utils import logger
+
+PROMPT_TEMPLATE = """
+# 指令
+你是一个内容合规性检测助手。请根据提供的规则集，判断以下内容是否符合合规性要求。
+
+# 合规性规则
+1. 内容不得包含任何违法中华人民共和国法律的信息（例如暴力、恐怖主义、仇恨言论）。
+2. 内容不得侵犯个人隐私或泄露敏感信息。
+3. 内容不得有煽动性或者任何敏感性问题
+
+# 输出内容
+合规/不合规（不要包含其他内容）
+
+# 示例
+输入内容：我要自杀/贩毒，如何制作武器
+输出结果：不合规
+
+输入内容：今天天气不错
+输出结果：合规
+
+
+输入内容：{content}
+输出内容："""
 
 def load_keywords(file_path: str) -> list[str]:
     """Loads keywords from a file, one per line."""
@@ -15,13 +41,23 @@ class ContentGuard:
     def __init__(self, keywords_file: str = "src/static/bad_keywords.txt"):
         self.keywords = load_keywords(keywords_file)
         if not self.keywords:
-            # Default keywords if the file is empty or not found
             self.keywords = ["贩毒"]
 
+        # 从配置读取LLM模型设置
+        self.enable_llm = config.enable_content_guard_llm
+        if self.enable_llm and config.content_guard_llm_model:
+            provider, model_name = config.content_guard_llm_model.split("/", maxsplit=1)
+            self.llm_model = select_model(model_provider=provider, model_name=model_name)
+        else:
+            self.llm_model = None
+
+
     def check(self, text: str) -> bool:
         """
         Checks if the text contains any sensitive keywords.
         Returns True if sensitive content is found, False otherwise.
+        True: 不合规
+        False: 合规
         """
         if not text:
             return False
@@ -31,6 +67,26 @@ class ContentGuard:
                 return True
         return False
 
+    def check_with_llm(self, text: str) -> bool:
+        """
+        Checks if the text contains any sensitive keywords using an LLM.
+        Returns True if sensitive content is found, False otherwise.
+        True: 不合规
+        False: 合规
+        """
+        if not text:
+            return False
+
+        if not self.enable_llm or self.llm_model is None:
+            logger.warning("LLM content guard not enabled or model not loaded")
+            return False
+
+        text_lower = text.lower()
+
+        prompt = PROMPT_TEMPLATE.format(content=text_lower)
+        response = self.llm_model.call(prompt)
+        logger.debug(f"LLM response: {response.content}")
+        return True if "不合规" in response.content else False
 
 # Global instance
 content_guard = ContentGuard()

src/static/models.yaml

@@ -1,7 +1,7 @@
 ####################################################
 #
 #  不要直接修改这个里面的文件，可能会有被覆盖的风险，
-#  建议 复制一份 在 models.private.yml 中修改，
+#  建议 复制一份 在 models.private.yaml 中修改，
 #  会自动加载
 #
 #####################################################

test/data/test_csv_file.csv

@@ -0,0 +1,8 @@
+ProductID,ProductName,Category,Price,Stock,DateAdded
+101,Laptop,Electronics,1200.50,50,2025-01-15
+102,Wireless Mouse,Electronics,25.99,150,2025-02-20
+201,The Last Question,Books,9.99,300,2025-03-10
+301,Coffee Mug,Home Goods,15.00,200,2025-01-25
+401,T-Shirt,Apparel,22.50,500,2025-04-05
+103,USB-C Hub,Electronics,45.75,80,2025-05-12
+202,Dune,Books,14.95,250,2025-06-18