敏感词拦截

dbgpt/app/openapi/api_v1/api_v1.py

@@ -472,7 +472,7 @@ async def no_stream_generator(chat):
         msg = await chat.nostream_call()
         yield f"data: {msg}\n\n"
 
-async def stream_call_sensitive_run_one(self):
+async def stream_call_sensitive_run_one():
     """
     只运行一次
     判断出用户输入有敏感词的时候,不需要调用大模型
@@ -489,7 +489,7 @@ async def stream_generator_sensitive(
     当用户输入有敏感词的时候,统一回答
     """
     span = root_tracer.start_span("stream_generator_sensitive")
-    async for _ in stream_call_sensitive_run_one(1):
+    async for _ in stream_call_sensitive_run_one():
         if incremental:
             yield f"data: {dfa_result}\n\n"
         else:

dbgpt/app/openapi/api_v2.py

@@ -16,6 +16,7 @@ from dbgpt.app.openapi.api_v1.api_v1 import (
     get_chat_instance,
     get_executor,
     stream_generator,
+    stream_generator_sensitive,
 )
 from dbgpt.app.scene import BaseChat, ChatScene
 from dbgpt.client.schema import ChatCompletionRequestBody, ChatMode
@@ -38,6 +39,10 @@ from dbgpt.serve.flow.service.service import Service as FlowService
 from dbgpt.util.executor_utils import blocking_func_to_async
 from dbgpt.util.tracer import SpanType, root_tracer
 
+from dbgpt.app.apps.utils.filter import mydfafiter
+from dbgpt.app.apps.vadmin.auth.utils.current import AllUserAuth, FullAdminAuth, OpenAuth
+from dbgpt.app.apps.vadmin.auth.utils.validation.auth import Auth
+
 router = APIRouter()
 api_settings = APISettings()
 get_bearer_token = HTTPBearer(auto_error=False)
@@ -74,6 +79,7 @@ async def check_api_key(
 @router.post("/v2/chat/completions", dependencies=[Depends(check_api_key)])
 async def chat_completions(
     request: ChatCompletionRequestBody = Body(),
+    auth: Auth = Depends(OpenAuth()),
 ):
     """Chat V2 completions
     Args:
@@ -91,6 +97,24 @@ async def chat_completions(
         "Connection": "keep-alive",
         "Transfer-Encoding": "chunked",
     }
+
+    # 从数据库中加载 并且初始化敏感词-->到内存中
+    await mydfafiter.parse_from_db(auth.db)
+
+    # 先判断敏感词
+    dfa_result, is_sensitive, matched_sensitives = mydfafiter.filter(request.user_input, "*")
+    print(f"--->敏感词替换后的问题:{dfa_result}")
+    print(f"--->匹配到的敏感词列表:{matched_sensitives}")
+
+    if is_sensitive:
+        print('用户输入有敏感词')
+        result = f"非常抱歉，您的提问中包含了部分敏感词汇:{matched_sensitives}，我作为AI，在回答时需遵守相关法律法规和政策，我们系统对特定敏感内容进行了限制。请您尝试调整或重新组织您的问题，避开这些敏感点，我将非常乐意为您提供准确且有帮助的回答。"
+        return StreamingResponse(
+            stream_generator_sensitive(request.incremental, result),
+            headers=headers,
+            media_type="text/plain",
+        )
+
     # check chat request
     check_chat_request(request)
     if request.conv_uid is None: