兼容思考模型

rag2_0/intent_recognition/IntentRecognition.py

@@ -155,7 +155,9 @@ class AsyncIntentRecognizer:
             logging.info(f"异步意图分类耗时统计 - 总耗时: {classification_time:.2f}秒")
 
             # 尝试直接解析JSON响应
-            parsed_output = classification_parser.parse(response.content.strip())
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = classification_parser.parse(clean_output)
             return parsed_output
         except Exception as e:
             raise RuntimeError(f"解析分类结果时出错: {e}") from e
@@ -216,7 +218,9 @@ class AsyncIntentRecognizer:
             response = await self._llm.invoke_async(formatted_prompt, False)
             
             # 尝试使用Pydantic解析器解析TermList
-            parsed_output = terms_list_parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = terms_list_parser.parse(clean_output)
             return parsed_output.terms
 
     
@@ -344,9 +348,9 @@ class AsyncIntentRecognizer:
         try:
             # 异步调用LLM
             response = await self._llm.invoke_async(formatted_prompt, False)
-        
+            response.content = response.content.strip()
-            # 尝试直接解析JSON响应
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
-            parsed_output = query_rewrite_parser.parse(response.content)
+            parsed_output = query_rewrite_parser.parse(clean_output)
             rewrite_end_time = time.time()
             rewrite_time = rewrite_end_time - rewrite_start_time
             logging.info(f"异步问题改写耗时统计 - 总耗时: {rewrite_time:.2f}秒")
@@ -597,9 +601,10 @@ class AsyncIntentRecognizer:
         try:
             # 异步调用LLM
             response = await self._llm.invoke_async(formatted_prompt, False)
-        
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
             # 尝试解析LLM响应
-            parsed_output = slot_parser.parse(response.content)
+            parsed_output = slot_parser.parse(clean_output)
             return parsed_output
         except Exception as e:
             # 如果解析失败，创建一个空的模型实例
@@ -633,7 +638,9 @@ class AsyncIntentRecognizer:
             response = await self._llm.invoke_async(formatted_prompt, False)
             
             # 解析输出
-            parsed_output = step_back_parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = step_back_parser.parse(clean_output)
             step_back_end_time = time.time()
             step_back_time = step_back_end_time - step_back_start_time
             logging.debug(f"异步后退提示生成耗时统计 - 总耗时: {step_back_time:.2f}秒")
@@ -670,7 +677,9 @@ class AsyncIntentRecognizer:
             response = await self._llm.invoke_async(formatted_prompt, False)
             
             # 解析输出
-            parsed_output = follow_up_parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = follow_up_parser.parse(clean_output)
             follow_up_end_time = time.time()
             follow_up_time = follow_up_end_time - follow_up_start_time
             logging.debug(f"异步后续问题生成耗时统计 - 总耗时: {follow_up_time:.2f}秒")
@@ -707,7 +716,9 @@ class AsyncIntentRecognizer:
             response = await self._llm.invoke_async(formatted_prompt, False)
             
             # 解析输出
-            parsed_output = hyde_parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = hyde_parser.parse(clean_output)
             hyde_end_time = time.time()
             hyde_time = hyde_end_time - hyde_start_time
             logging.debug(f"异步假设性文档生成耗时统计 - 总耗时: {hyde_time:.2f}秒")
@@ -744,7 +755,9 @@ class AsyncIntentRecognizer:
             response = await self._llm.invoke_async(formatted_prompt, False)
             
             # 解析输出
-            parsed_output = multi_questions_parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            parsed_output = multi_questions_parser.parse(clean_output)
             multi_questions_end_time = time.time()
             multi_questions_time = multi_questions_end_time - multi_questions_start_time
             logging.debug(f"异步多角度问题生成耗时统计 - 总耗时: {multi_questions_time:.2f}秒")
@@ -799,7 +812,9 @@ class AsyncIntentRecognizer:
 
         try:
             # 解析LLM响应为JSON
-            result_json = parser.parse(response.content)
+            response.content = response.content.strip()
+            clean_output = re.sub(r'<think>.*?</think>', '', response.content, flags=re.DOTALL)
+            result_json = parser.parse(clean_output)
             classification = result_json.classification
             slot_filling = result_json.slots
             is_complete, missing_slots = slot_filling.check_required_slots()