Update app.py

app.py

@@ -103,6 +103,8 @@ class TextToSQLSystem:
         self.log_history = []
         self._log("初始化系統...")
         self.query_cache = {}
+        self.backend = None  # 'gguf' | 'transformers' | None
+        self.gguf_llm = None  # 實際 llama.cpp 物件
 
         # 1. 載入嵌入模型
         self._log(f"載入嵌入模型: {embed_model_name}")
@@ -149,7 +151,7 @@ class TextToSQLSystem:
             )
 
             # 使用一組更基礎、更穩定的參數來載入模型
-            self.llm = Llama(
+            self.gguf_llm = Llama(
                 model_path=model_path,
                 n_ctx=2048,      # 將上下文增加到 2048 以確保 Prompt 不會超長
                 n_threads=4,     # 保持 4 線程
@@ -159,7 +161,8 @@ class TextToSQLSystem:
             )
 
             # 簡單測試模型是否能回應
-            self.llm("你好", max_tokens=3)
+            self.gguf_llm("你好", max_tokens=3)
+            self.backend = "gguf"
             self._log("✅ GGUF 模型載入成功")
 
         except Exception as e:
@@ -176,7 +179,7 @@ class TextToSQLSystem:
                 repo_type="dataset"
             )
 
-            self.llm = Llama(
+            self.gguf_llm = Llama(
                 model_path=model_path,
                 n_ctx=512,
                 n_threads=4,
@@ -185,7 +188,7 @@ class TextToSQLSystem:
             )
 
             # 測試生成
-            test_result = self.llm("SELECT", max_tokens=5)
+            test_result = self.gguf_llm("SELECT", max_tokens=5)
             self._log("✅ GGUF 模型載入成功")
             return True
 
@@ -223,63 +226,96 @@ class TextToSQLSystem:
                 pad_token_id=self.transformers_tokenizer.eos_token_id
             )
 
-            self.llm = "transformers"  # 標記使用 transformers
+            # 標記目前後端為 transformers
+            self.backend = "transformers"
             self._log("✅ Transformers 模型載入成功")
 
         except Exception as e:
             self._log(f"❌ Transformers 載入也失敗: {e}", "ERROR")
-            self.llm = None
 
     def huggingface_api_call(self, prompt: str) -> str:
-        """調用 GGUF 模型，並加入詳細的原始輸出日誌"""
-        if self.llm is None:
-            self._log("模型未載入，返回 fallback SQL。", "ERROR")
-            return self._generate_fallback_sql(prompt)
-
-        try:
-            # 重要: 移除 ";" 讓模型可輸出完整查詢（包含結尾分號前所有內容）
-            output = self.llm(
-                prompt,
-                max_tokens=350,
-                temperature=0.05,
-                top_p=0.9,
-                echo=False,
-                stop=["```", "\n\n", "</s>"]
-            )
-
-            self._log(f"🧠 模型原始輸出 (Raw Output): {output}", "DEBUG")
-
-
-            if output and "choices" in output and len(output["choices"]) > 0:
-                generated_text = output["choices"][0]["text"]
+        """生成 SQL：優先使用 transformers，其次 gguf，最後 fallback"""
+        # transformers 後端
+        if self.backend == "transformers" and hasattr(self, "generation_pipeline"):
+            try:
+                gen = self.generation_pipeline(
+                    prompt,
+                    max_new_tokens=350,
+                    do_sample=True,
+                    temperature=0.05,
+                    top_p=0.9
+                )
+                # 盡量從 pipeline 結果提取文字
+                generated_text = ""
+                try:
+                    if isinstance(gen, list) and gen:
+                        first = gen[0]
+                        if isinstance(first, dict) and "generated_text" in first:
+                            generated_text = str(first["generated_text"])  # type: ignore[index]
+                        else:
+                            generated_text = str(first)
+                    else:
+                        generated_text = str(gen)
+                except Exception:
+                    generated_text = str(gen)
+                # 若包含 prompt，裁切前綴
+                if isinstance(generated_text, str) and generated_text.startswith(prompt):
+                    generated_text = generated_text[len(prompt):]
                 self._log(f"📝 提取出的生成文本: {generated_text.strip()}", "DEBUG")
 
-                # --- 新增的清理邏輯 ---
                 lines = generated_text.strip().split('\n')
-                # 過濾掉所有以 '--' 開頭的註解行
                 non_comment_lines = [line for line in lines if not line.strip().startswith('--')]
                 cleaned_text = "\n".join(non_comment_lines).strip()
-
                 if cleaned_text != generated_text.strip():
                     self._log(f"🧹 清理掉註解後的文本: {cleaned_text}", "DEBUG")
+                if cleaned_text and not re.match(r"^\s*select\b", cleaned_text, flags=re.IGNORECASE):
+                    self._log("⚙️ 補上缺失的 'SELECT ' 起手以形成完整查詢", "DEBUG")
+                    cleaned_text = "SELECT " + cleaned_text.lstrip()
+                return cleaned_text
+            except Exception as e:
+                self._log(f"❌ Transformers 生成失敗: {e}", "ERROR")
+                return ""
 
-                return cleaned_text  # <-- 返回清理後的文本
-                # --- 清理邏輯結束 ---
-            else:
-                self._log("❌ 模型的原始輸出格式不正確或為空。", "ERROR")
+        # gguf 後端
+        if self.backend == "gguf" and self.gguf_llm is not None and callable(getattr(self.gguf_llm, "__call__", None)):
+            try:
+                output = self.gguf_llm(
+                    prompt,
+                    max_tokens=350,
+                    temperature=0.05,
+                    top_p=0.9,
+                    echo=False,
+                    stop=["```"]
+                )
+                self._log(f"🧠 模型原始輸出 (Raw Output): {output}", "DEBUG")
+                if output and "choices" in output and len(output["choices"]) > 0:
+                    generated_text = output["choices"][0]["text"]
+                    self._log(f"📝 提取出的生成文本: {generated_text.strip()}", "DEBUG")
+                    lines = str(generated_text).strip().split('\n')
+                    non_comment_lines = [line for line in lines if not line.strip().startswith('--')]
+                    cleaned_text = "\n".join(non_comment_lines).strip()
+                    if cleaned_text != str(generated_text).strip():
+                        self._log(f"🧹 清理掉註解後的文本: {cleaned_text}", "DEBUG")
+                    if cleaned_text and not re.match(r"^\s*select\b", cleaned_text, flags=re.IGNORECASE):
+                        self._log("⚙️ 補上缺失的 'SELECT ' 起手以形成完整查詢", "DEBUG")
+                        cleaned_text = "SELECT " + cleaned_text.lstrip()
+                    return cleaned_text
+                else:
+                    self._log("❌ 模型的原始輸出格式不正確或為空。", "ERROR")
+                    return ""
+            except Exception as e:
+                self._log(f"❌ GGUF 生成失敗: {e}", "ERROR")
                 return ""
 
-        except Exception as e:
-            self._log(f"❌ 模型生成過程中發生嚴重錯誤: {e}", "CRITICAL")
-            import traceback
-            self._log(traceback.format_exc(), "DEBUG")
-            return ""
+        # 後備：都不可用時，回退
+        self._log("模型未載入或不可用，返回 fallback SQL。", "ERROR")
+        return self._generate_fallback_sql(prompt)
 
     def _load_gguf_model_fallback(self, model_path):
         """備用載入方式"""
         try:
             # 嘗試不同的參數組合
-            self.llm = Llama(
+            self.gguf_llm = Llama(
                 model_path=model_path,
                 n_ctx=512,  # 更小的上下文
                 n_threads=4,
@@ -292,7 +328,7 @@ class TextToSQLSystem:
             self._log("✅ 備用方式載入成功")
         except Exception as e:
             self._log(f"❌ 備用方式也失敗: {e}", "ERROR")
-            self.llm = None
+            self.gguf_llm = None
 
     def _log(self, message: str, level: str = "INFO"):
         self.log_history.append(format_log(message, level))
@@ -548,7 +584,7 @@ class TextToSQLSystem:
                 }
                 break
 
-        # ==============================================================================
+    # ==============================================================================
         #  第一層：模組化意圖偵測與動態SQL組合
         # ==============================================================================
 
@@ -622,6 +658,55 @@ class TextToSQLSystem:
             sql_components['where'].append(f"jip.LabGroup = '{db_lab_group}'")
             sql_components['log_parts'].append(f"{user_input_group}組(->{db_lab_group})")
 
+        # --- 2.6: 兩年份比較模板（優先級：高） ---
+        # 偵測『比較/vs/對比/相較/相比』字樣，擷取兩個年份與（可選）買家名稱
+        compare_hit = any(kw in q_lower for kw in ["比較", "對比", "相較", "相比", "vs", "versus"])
+        years_found = re.findall(r"(20\d{2})", question)
+        years_unique = []
+        for y in years_found:
+            if y not in years_unique:
+                years_unique.append(y)
+        if compare_hit and len(years_unique) >= 2:
+            year_a, year_b = years_unique[0], years_unique[1]
+            # 嘗試抓買家名稱（英文/數字/符號），若沒有則不加 buyer 條件
+            buyer_name = None
+            buyer_match = re.search(r"(?:買家|买家|buyer)\s*[:：]?\s*([A-Za-z0-9&.\- ]+)", question, re.IGNORECASE)
+            if buyer_match:
+                buyer_name = buyer_match.group(1).strip()
+
+            # 判斷偏向金額或件數
+            amount_intent = any(kw in q_lower for kw in ["金額", "金钱", "amount", "營收", "業績", "營業��", "銷售額", "revenue"])
+
+            if amount_intent:
+                # 金額版：需要發票表，依架構命名使用 TSR53Invoice 與 LocalAmount；與樣本描述以 JobNo 關聯
+                sql = (
+                    "SELECT strftime('%Y', jt.ReportAuthorization) AS year, "
+                    "SUM(COALESCE(inv.LocalAmount, 0)) AS total_amount "
+                    "FROM JobTimeline AS jt "
+                    "JOIN TSR53SampleDescription AS sd ON sd.JobNo = jt.JobNo "
+                    "LEFT JOIN TSR53Invoice AS inv ON inv.JobNo = jt.JobNo "
+                    "WHERE jt.ReportAuthorization IS NOT NULL "
+                    f"AND strftime('%Y', jt.ReportAuthorization) IN ('{year_a}', '{year_b}') "
+                )
+                if buyer_name:
+                    sql += f"AND sd.BuyerName LIKE '%{buyer_name}%' "
+                sql += "GROUP BY year ORDER BY year;"
+                return self._finalize_sql(sql, f"模板覆寫: 兩年份金額比較 {year_a} vs {year_b}" )
+            else:
+                # 件數版：以報告數量為主，去重 JobNo
+                sql = (
+                    "SELECT strftime('%Y', jt.ReportAuthorization) AS year, "
+                    "COUNT(DISTINCT jt.JobNo) AS report_count "
+                    "FROM JobTimeline AS jt "
+                    "JOIN TSR53SampleDescription AS sd ON sd.JobNo = jt.JobNo "
+                    "WHERE jt.ReportAuthorization IS NOT NULL "
+                    f"AND strftime('%Y', jt.ReportAuthorization) IN ('{year_a}', '{year_b}') "
+                )
+                if buyer_name:
+                    sql += f"AND sd.BuyerName LIKE '%{buyer_name}%' "
+                sql += "GROUP BY year ORDER BY year;"
+                return self._finalize_sql(sql, f"模板覆寫: 兩年份件數比較 {year_a} vs {year_b}" )
+
         # --- 3. 判斷是否觸發了模板，並動態組合 SQL ---
         if 'action' in intents:
             sql_components['from'] = "FROM JobTimeline AS jt"
@@ -765,6 +850,7 @@ class TextToSQLSystem:
 User question: "{user_q}"
 Your single SQLite query response:
 ```sql
+SELECT
 """
         self._log(f"📏 Prompt 長度: {len(prompt)} 字符")
         # 不再需要複雜的長度截斷邏輯，因為 schema 已經被簡化
@@ -842,7 +928,7 @@ Your single SQLite query response:
 
             # --- 新增：如果是第二次嘗試，加入修正指令 ---
             if attempt > 0:
-                correction_prompt = "\nYour previous attempt failed because you did not provide a valid SQL query. REMEMBER: ONLY output the SQL code inside a ```sql block. DO NOT write comments or explanations.\nSQL:\n```sql\n"
+                correction_prompt = "\nYour previous attempt failed because you did not provide a valid SQL query. REMEMBER: ONLY output the SQL code inside a ```sql block. DO NOT write comments or explanations.\nSQL:\n```sql\nSELECT "
                 # 將原本 prompt 的結尾替換成我們的修正指令
                 prompt = prompt.rsplit("SQL:\n```sql", 1)[0] + correction_prompt