Update app.py

app.py

@@ -147,16 +147,21 @@ class TextToSQLSystem:
         # 檢查內存狀況
         self._log(check_memory_usage())
 
-        # 1. 載入嵌入模型
-        self._log(f"載入嵌入模型: {embed_model_name}")
-        self.embed_tokenizer = AutoTokenizer.from_pretrained(embed_model_name)
-        self.embed_model = AutoModel.from_pretrained(embed_model_name)
-        try:
-            self.embed_model.to(self.embed_device)
-            self._log(f"嵌入模型設備: {self.embed_device}")
-        except Exception as e:
-            self._log(f"將嵌入模型移動到設備失敗: {e}", "WARNING")
-            self.embed_device = "cpu"
+        # 1. 嵌入模型（在禁用索引時略過以節省記憶體）
+        if ENABLE_INDEX:
+            self._log(f"載入嵌入模型: {embed_model_name}")
+            self.embed_tokenizer = AutoTokenizer.from_pretrained(embed_model_name)
+            self.embed_model = AutoModel.from_pretrained(embed_model_name)
+            try:
+                self.embed_model.to(self.embed_device)
+                self._log(f"嵌入模型設備: {self.embed_device}")
+            except Exception as e:
+                self._log(f"將嵌入模型移動到設備失敗: {e}", "WARNING")
+                self.embed_device = "cpu"
+        else:
+            self.embed_tokenizer = None
+            self.embed_model = None
+            self._log("ENABLE_INDEX=0，略過嵌入模型載入以節省記憶體")
 
         # 2. 載入數據庫結構
         self.schema = self._load_schema()
@@ -263,7 +268,8 @@ class TextToSQLSystem:
                 temperature=0.1,
                 top_p=0.9,
                 echo=False,
-                stop=["```", ";", "\n\n", "</s>"],
+                # 避免在分號處截斷
+                stop=["```", "\n\n", "</s>"],
             )
             elapsed = (datetime.now() - start_ts).total_seconds()
             self._log(f"推論耗時: {elapsed:.2f}s", "DEBUG")
@@ -311,6 +317,10 @@ class TextToSQLSystem:
         """編碼文本為嵌入向量"""
         if isinstance(texts, str):
             texts = [texts]
+        if (self.embed_model is None) or (self.embed_tokenizer is None):
+            # 在禁用索引情況下不應被呼叫；保險處理
+            self._log("嵌入模型未載入（ENABLE_INDEX=0），_encode_texts 被略過。", "WARNING")
+            return torch.empty((len(texts), 384))  # 回傳空張量佔位
         inputs = self.embed_tokenizer(texts, padding=True, truncation=True,
                                       return_tensors="pt", max_length=512)
         # 移動到對應設備
@@ -641,6 +651,19 @@ SELECT
             'group_by': [], 'order_by': [], 'log_parts': []
         }
 
+        # 先處理多年份比較：如 "2021 與 2022 比較"、"2021年跟2022年對比"
+        years = re.findall(r"(20\d{2})\s*年?", q)
+        is_compare = re.search(r"比較|對比|對照|compare|versus|vs\.?", q)
+        if len(set(years)) >= 2 and is_compare:
+            ys = sorted(set(years))[:4]
+            want_items = ("測試項目" in q) or ("item" in q_lower)
+            select_expr = "COUNT(jip.ItemCode) AS item_count" if want_items else "COUNT(DISTINCT jt.JobNo) AS report_count"
+            join_items = "JOIN JobItemsInProgress AS jip ON jt.JobNo = jip.JobNo" if want_items else ""
+            where_years = " OR ".join([f"strftime('%Y', jt.ReportAuthorization) = '{y}'" for y in ys])
+            template = f"SELECT strftime('%Y', jt.ReportAuthorization) AS 年份, {select_expr} FROM JobTimeline AS jt {join_items} WHERE jt.ReportAuthorization IS NOT NULL AND ({where_years}) GROUP BY 年份 ORDER BY 年份;"
+            self._log(f"🔄 多年份比較模板: years={','.join(ys)} items={want_items}")
+            return self._finalize_sql(template, f"模板覆寫: {','.join(ys)} 年比較")
+
         # 動作意圖：count / list
         if any(kw in q_lower for kw in ['幾份', '份數', '份数', '多少', '數量', '總數', 'how many', 'count']):
             intents['action'] = 'count'