Update query_understanding.py

query_understanding.py

@@ -43,11 +43,12 @@ class QueryUnderstandingEngine:
     def __init__(self):
         """初始化查詢理解引擎"""
         self.sbert_model = None
+        self._sbert_loading_attempted = False
         self.breed_list = self._load_breed_list()
         self.synonyms = self._initialize_synonyms()
         self.semantic_templates = {}
-        self._initialize_sbert_model()
-        self._build_semantic_templates()
+        # 延遲SBERT載入直到需要時才在GPU環境中進行
+        print("QueryUnderstandingEngine initialized (SBERT loading deferred)")
 
     def _load_breed_list(self) -> List[str]:
         """載入品種清單"""
@@ -66,25 +67,35 @@ class QueryUnderstandingEngine:
                    'Bulldog', 'Poodle', 'Beagle', 'Border_Collie', 'Yorkshire_Terrier']
 
     def _initialize_sbert_model(self):
-        """初始化 SBERT 模型"""
+        """初始化 SBERT 模型 - 延遲載入以避免ZeroGPU CUDA初始化問題"""
+        if self.sbert_model is not None or getattr(self, '_sbert_loading_attempted', False):
+            return self.sbert_model
+            
         try:
+            print("Loading SBERT model for query understanding in GPU context...")
             model_options = ['all-MiniLM-L6-v2', 'all-mpnet-base-v2', 'all-MiniLM-L12-v2']
 
             for model_name in model_options:
                 try:
-                    self.sbert_model = SentenceTransformer(model_name)
-                    print(f"SBERT model {model_name} loaded successfully for query understanding")
-                    return
+                    import torch
+                    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+                    self.sbert_model = SentenceTransformer(model_name, device=device)
+                    print(f"SBERT model {model_name} loaded successfully for query understanding on {device}")
+                    return self.sbert_model
                 except Exception as e:
                     print(f"Failed to load {model_name}: {str(e)}")
                     continue
 
             print("All SBERT models failed to load. Using keyword-only analysis.")
             self.sbert_model = None
+            return None
 
         except Exception as e:
             print(f"Failed to initialize SBERT model: {str(e)}")
             self.sbert_model = None
+            return None
+        finally:
+            self._sbert_loading_attempted = True
 
     def _initialize_synonyms(self) -> DimensionalSynonyms:
         """初始化多維度同義詞字典"""
@@ -143,6 +154,10 @@ class QueryUnderstandingEngine:
 
     def _build_semantic_templates(self):
         """建立語義模板向量（僅在 SBERT 可用時）"""
+        # Initialize SBERT model if needed
+        if self.sbert_model is None:
+            self._initialize_sbert_model()
+            
         if not self.sbert_model:
             return
 
@@ -192,6 +207,9 @@ class QueryUnderstandingEngine:
             dimensions = self._extract_keyword_dimensions(normalized_input)
 
             # 如果 SBERT 可用，進行語義分析增強
+            if self.sbert_model is None:
+                self._initialize_sbert_model()
+                
             if self.sbert_model:
                 semantic_dimensions = self._extract_semantic_dimensions(user_input)
                 dimensions = self._merge_dimensions(dimensions, semantic_dimensions)
@@ -435,7 +453,6 @@ class QueryUnderstandingEngine:
             ])
         }
 
-# 便利函數
 def analyze_user_query(user_input: str) -> QueryDimensions:
     """
     便利函數：分析使用者查詢
@@ -461,4 +478,4 @@ def get_query_summary(user_input: str) -> Dict[str, Any]:
     """
     engine = QueryUnderstandingEngine()
     dimensions = engine.analyze_query(user_input)
-    return engine.get_dimension_summary(dimensions)
+    return engine.get_dimension_summary(dimensions)