Update app.py

app.py

@@ -16,7 +16,7 @@ DATASET_REPO_ID = "Paul720810/Text-to-SQL-Softline"
 GGUF_REPO_ID = "Paul720810/gguf-models"
 GGUF_FILENAME = "qwen2.5-coder-1.5b-sql-finetuned.q4_k_m.gguf"
 
-FEW_SHOT_EXAMPLES_COUNT = 2
+FEW_SHOT_EXAMPLES_COUNT = 1
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
 print("=" * 60)
@@ -33,12 +33,19 @@ def format_log(message: str, level: str = "INFO") -> str:
     return f"[{get_current_time()}] [{level.upper()}] {message}"
 
 def parse_sql_from_response(response_text: str) -> Optional[str]:
-    """從模型輸出提取 SQL"""
+    # 1. 先找 ```sql ... ```
     match = re.search(r"```sql\n(.*?)\n```", response_text, re.DOTALL)
     if match:
         return match.group(1).strip()
+
+    # 2. 如果沒找到，嘗試找最長的 SELECT … 語句
+    match = re.search(r"(SELECT .*?;)", response_text, re.DOTALL | re.IGNORECASE)
+    if match:
+        return match.group(1).strip()
+
     return None
 
+
 # ==================== Text-to-SQL 核心類 ====================
 class TextToSQLSystem:
     def __init__(self, embed_model='sentence-transformers/paraphrase-multilingual-mpnet-base-v2'):
@@ -60,8 +67,9 @@ class TextToSQLSystem:
         )
         self.llm = Llama(
             model_path=model_path,
-            n_ctx=4096,
-            n_threads=os.cpu_count(),
+            n_ctx=2048,              # 少一半上下文 → 快
+            n_threads=os.cpu_count(),# 用滿 CPU
+            n_batch=512,             # 增加 batch size → 減少迭代
             verbose=False
         )
         self._log(f"✅ 已載入 GGUF 模型: {GGUF_FILENAME}")
@@ -138,7 +146,7 @@ class TextToSQLSystem:
             self._log("🧠 使用 GGUF 模型生成 SQL...")
             output = self.llm(
                 prompt,
-                max_tokens=256,
+                max_tokens=128,
                 temperature=0.2,
                 stop=["</s>", "```"],
                 echo=False