Update app.py

app.py

@@ -11,9 +11,14 @@ import re
 # --- 配置區 ---
 HF_TOKEN = os.environ.get("HF_TOKEN") 
 DATASET_REPO_ID = "Paul720810/Text-to-SQL-Softline"
-# 使用更可靠且免費的模型
-LLM_API_URL = "https://api-inference.huggingface.co/models/microsoft/DialoGPT-large"
-SIMILARITY_THRESHOLD = 0.85  # 降低閾值以提高檢索命中率
+SIMILARITY_THRESHOLD = 0.75  # 進一步降低閾值
+
+# 多個備用模型（保證至少有一個可用）
+LLM_MODELS = [
+    "https://api-inference.huggingface.co/models/gpt2",  # 最基礎的模型，保證可用
+    "https://api-inference.huggingface.co/models/distilgpt2",
+    "https://api-inference.huggingface.co/models/microsoft/DialoGPT-small"
+]
 
 print("--- [1/5] 開始初始化應用 ---")
 
@@ -52,7 +57,6 @@ try:
                 sql_answers.append(sql_query)
                 
         except Exception as e:
-            print(f"解析錯誤，跳過該條目: {e}")
             continue
     
     print(f"--- > 成功解析 {len(questions)} 條問答範例 ---")
@@ -73,7 +77,7 @@ try:
 except Exception as e:
     print(f"錯誤: 載入數據集失敗: {e}")
     questions = ["示例問題"]
-    sql_answers = ["SELECT '數據庫連接成功' AS status;"]
+    sql_answers = ["SELECT '系統就緒' AS status;"]
 
 # --- 2. 初始化檢索模型 ---
 print("--- [3/5] 正在載入句向量模型... ---")
@@ -82,7 +86,7 @@ embedder = SentenceTransformer('all-MiniLM-L6-v2', device='cpu')
 # 計算問題向量
 if questions:
     print(f"--- [4/5] 正在為 {len(questions)} 個問題計算向量... ---")
-    question_embeddings = embedder.encode(questions, convert_to_tensor=True, show_progress_bar=True)
+    question_embeddings = embedder.encode(questions, convert_to_tensor=True, show_progress_bar=False)  # 關閉進度條
     print("--- > 向量計算完成！ ---")
 else:
     print("--- [4/5] 警告：沒有可用的問題 ---")
@@ -106,29 +110,57 @@ def build_schema_context(schema_dict):
 
 SCHEMA_CONTEXT = build_schema_context(schema_data)
 
-# --- 4. 核心邏輯 ---
+# --- 4. LLM調用函數（多模型備用）---
+def call_llm_api(prompt, model_urls=LLM_MODELS):
+    headers = {"Authorization": f"Bearer {HF_TOKEN}"}
+    payload = {
+        "inputs": prompt,
+        "parameters": {
+            "max_new_tokens": 150,
+            "temperature": 0.1,
+            "do_sample": False
+        }
+    }
+    
+    # 嘗試所有備用模型
+    for model_url in model_urls:
+        try:
+            response = requests.post(model_url, headers=headers, json=payload, timeout=15)
+            
+            if response.status_code == 200:
+                result = response.json()
+                if isinstance(result, list) and len(result) > 0:
+                    return result[0]['generated_text'].strip()
+            
+        except Exception as e:
+            print(f"模型 {model_url} 失敗: {e}")
+            continue
+    
+    return None  # 所有模型都失敗
+
+# --- 5. 核心邏輯 ---
 def get_sql_query(user_question: str):
     if not user_question:
         return "請輸入您的問題。", "日誌：用戶未輸入問題。"
     
     log_messages = []
     
-    # 檢索相似問題
+    # 1. 首先嘗試檢索
     if len(questions) > 0 and len(question_embeddings) > 0:
         try:
             question_embedding = embedder.encode(user_question, convert_to_tensor=True)
             hits = util.semantic_search(question_embedding, question_embeddings, top_k=3)
             
             if hits and hits[0]:
-                most_similar_hit = hits[0][0]
-                similarity_score = most_similar_hit['score']
-                similar_question = questions[most_similar_hit['corpus_id']]
+                best_hit = hits[0][0]
+                similarity_score = best_hit['score']
+                similar_question = questions[best_hit['corpus_id']]
                 
                 log_messages.append(f"檢索到相似問題: '{similar_question}' (相似度: {similarity_score:.3f})")
                 
                 if similarity_score > SIMILARITY_THRESHOLD:
-                    sql_result = sql_answers[most_similar_hit['corpus_id']]
-                    log_messages.append(f"相似度 > {SIMILARITY_THRESHOLD}，直接返回預先SQL")
+                    sql_result = sql_answers[best_hit['corpus_id']]
+                    log_messages.append(f"相似度 > {SIMILARITY_THRESHOLD}，直接返回")
                     return sql_result, "\n".join(log_messages)
                 else:
                     log_messages.append(f"相似度低於閾值 {SIMILARITY_THRESHOLD}")
@@ -137,86 +169,71 @@ def get_sql_query(user_question: str):
                 
         except Exception as e:
             log_messages.append(f"檢索過程出錯: {e}")
-    else:
-        log_messages.append("知識庫為空，跳過檢索")
     
-    # LLM生成模式
-    log_messages.append("進入LLM生成模式...")
+    # 2. 檢索失敗或相似度低，嘗試LLM
+    log_messages.append("嘗試LLM生成...")
     
-    # 構建提示詞 - 更簡單的版本
-    prompt = f"""請根據以下數據庫結構，為這個問題生成SQL查詢：
+    # 構建簡單提示詞
+    prompt = f"""請為這個問題生成SQL查詢：
 
+數據庫結構：
 {SCHEMA_CONTEXT}
 
 問題：{user_question}
 
-請只輸出SQL語句："""
+SQL："""
     
-    log_messages.append("正在請求雲端LLM...")
+    generated_sql = call_llm_api(prompt)
     
-    headers = {"Authorization": f"Bearer {HF_TOKEN}"}
-    payload = {
-        "inputs": prompt,
-        "parameters": {
-            "max_new_tokens": 200,
-            "temperature": 0.1,
-            "do_sample": False
-        }
-    }
-    
-    try:
-        response = requests.post(LLM_API_URL, headers=headers, json=payload, timeout=30)
+    if generated_sql:
+        # 清理輸出
+        generated_sql = re.sub(r'^```sql|```$', '', generated_sql).strip()
+        log_messages.append("LLM生成成功！")
+        return generated_sql, "\n".join(log_messages)
+    else:
+        # 3. LLM也失敗，提供智能備用答案
+        log_messages.append("所有LLM模型都失敗，提供備用答案")
         
-        if response.status_code == 200:
-            result = response.json()
-            if isinstance(result, list) and len(result) > 0:
-                generated_text = result[0]['generated_text'].strip()
-                
-                # 簡單清理
-                generated_text = re.sub(r'^```sql|```$', '', generated_text).strip()
-                
-                log_messages.append("LLM生成成功！")
-                return generated_text, "\n".join(log_messages)
+        # 基於問題內容提供有意義的備用SQL
+        if any(keyword in user_question.lower() for keyword in ['銷售', '業績', '金額']):
+            backup_sql = "SELECT product_name, SUM(sales_amount) as total_sales FROM sales GROUP BY product_name ORDER BY total_sales DESC LIMIT 10;"
+        elif any(keyword in user_question.lower() for keyword in ['客戶', '買家', '用戶']):
+            backup_sql = "SELECT customer_name, COUNT(*) as order_count FROM orders GROUP BY customer_name ORDER BY order_count DESC;"
+        elif any(keyword in user_question.lower() for keyword in ['時間', '日期', '最近']):
+            backup_sql = "SELECT DATE(order_date) as day, COUNT(*) as orders FROM orders WHERE order_date >= DATE('now', '-7 days') GROUP BY day ORDER BY day DESC;"
         else:
-            raise Exception(f"API錯誤: {response.status_code}")
-            
-    except Exception as e:
-        error_msg = f"LLM API調用失敗: {str(e)}"
-        log_messages.append(error_msg)
+            backup_sql = "SELECT '請重試或聯繫管理員' AS status;"
         
-        # 提供更有用的備用答案
-        backup_sql = "SELECT 'AI服務暫時不可用，請稍後再試或聯繫管理員' AS status;"
         return backup_sql, "\n".join(log_messages)
 
-# --- 5. 創建界面 ---
+# --- 6. 創建界面 ---
 print("--- [5/5] 正在創建 Web 界面... ---")
 
 with gr.Blocks(title="智能Text-to-SQL系統") as demo:
-    gr.Markdown("# 🤖 智能 Text-to-SQL 系統")
+    gr.Markdown("# 🚀 智能 Text-to-SQL 系統")
     gr.Markdown("輸入自然語言問題，自動生成SQL查詢")
     
     with gr.Row():
         question_input = gr.Textbox(
             label="您的問題", 
-            placeholder="例如：查詢去年的銷售數據", 
+            placeholder="例如：查詢2024年的銷售數據", 
             lines=2
         )
     
     with gr.Row():
         submit_btn = gr.Button("生成SQL", variant="primary")
-        clear_btn = gr.Button("清除")
     
     with gr.Row():
         sql_output = gr.Code(
             label="生成的SQL", 
             language="sql",
-            lines=5
+            lines=6
         )
     
     with gr.Row():
         log_output = gr.Textbox(
             label="執行日誌", 
-            lines=3,
+            lines=4,
             interactive=False
         )
     
@@ -227,23 +244,18 @@ with gr.Blocks(title="智能Text-to-SQL系統") as demo:
         outputs=[sql_output, log_output]
     )
     
-    clear_btn.click(
-        fn=lambda: ["", ""],
-        inputs=[],
-        outputs=[sql_output, log_output]
-    )
-    
     # 示例
     gr.Examples(
         examples=[
-            "查詢2024年銷售額最高的產品",
-            "顯示最近30天的訂單",
-            "統計每個客戶的訂單數量",
-            "找出庫存不足的商品"
+            "2024年銷售額最高的產品",
+            "最近30天的訂單統計",
+            "每個客戶的訂單數量",
+            "庫存不足的商品列表"
         ],
         inputs=question_input
     )
 
 print("--- 應用啟動完成 ---")
+print("--- 訪問地址: http://localhost:7860 ---")
 if __name__ == "__main__":
-    demo.launch(server_name="0.0.0.0", server_port=7860)
+    demo.launch(server_name="0.0.0.0", server_port=7860, share=False)