Update app.py

app.py

@@ -8,6 +8,7 @@ import json
 import re
 import concurrent.futures
 import time
+import random
 from pdf2image import convert_from_path
 from PIL import Image
 from dotenv import load_dotenv
@@ -20,6 +21,7 @@ from pptx.dml.color import RGBColor
 # 使用 Google 新版 SDK
 from google import genai
 from google.genai import types
+from google.api_core import exceptions as google_exceptions
 
 load_dotenv()
 
@@ -48,12 +50,39 @@ class NotebookLMTool:
         except:
             return []
 
-    # --- 單頁處理邏輯 (獨立出來以便平行運算) ---
+    def _call_gemini_with_retry(self, model_name, contents, config=None, retries=5):
+        """
+        封裝 Gemini 呼叫，加入指數退避重試機制 (Exponential Backoff)
+        專門處理 429 Resource Exhausted 錯誤
+        """
+        delay = 10 # 初始等待秒數
+        
+        for attempt in range(retries):
+            try:
+                response = self.client.models.generate_content(
+                    model=model_name,
+                    contents=contents,
+                    config=config
+                )
+                return response
+            except Exception as e:
+                # 檢查是否為 Rate Limit 相關錯誤 (包含 429 或 Service Unavailable)
+                error_str = str(e)
+                if "429" in error_str or "RESOURCE_EXHAUSTED" in error_str or "503" in error_str:
+                    wait_time = delay + random.uniform(0, 5) # 加入隨機抖動避免同時重試
+                    print(f"⚠️ 觸發 API 限制，暫停 {wait_time:.1f} 秒後重試 ({attempt+1}/{retries})...", flush=True)
+                    time.sleep(wait_time)
+                    delay *= 2 # 等待時間加倍 (10s -> 20s -> 40s...)
+                else:
+                    raise e # 其他錯誤直接拋出
+        
+        raise Exception("API 重試次數過多，請稍後再試。")
+
+    # --- 單頁處理邏輯 ---
     def process_single_page(self, page_index, img, img_output_dir):
         """處理單一頁面的：去字(背景) + 文字分析(Layout)"""
         print(f"🚀 [Page {page_index+1}] 開始處理...", flush=True)
         
-        # 結果容器
         result = {
             "index": page_index,
             "bg_path": None,
@@ -78,9 +107,9 @@ class NotebookLMTool:
             3. Output ONLY the image.
             """
             
-            # 使用 2.0-flash-exp 進行繪圖
-            resp_img = self.client.models.generate_content(
-                model="gemini-2.5-flash-image", 
+            # 使用帶重試機制的呼叫
+            resp_img = self._call_gemini_with_retry(
+                model_name="gemini-2.0-flash-exp", 
                 contents=[clean_prompt, img],
                 config=types.GenerateContentConfig(response_modalities=["IMAGE"])
             )
@@ -113,9 +142,9 @@ class NotebookLMTool:
         # 失敗回退原圖
         if not bg_success:
             img.save(final_bg_path)
-            result["bg_path"] = final_bg_path # 仍需路徑給 PPT 使用
+            result["bg_path"] = final_bg_path 
             result["preview"] = (final_bg_path, f"Page {page_index+1} (Original)")
-            result["log"] += f"[P{page_index+1}] Warning: Background cleaning failed.\n"
+            result["log"] += f"[P{page_index+1}] Warning: Background cleaning failed (Rate Limit or Error).\n"
 
         # 2. 文字與佈局分析 (Layout Analysis)
         try:
@@ -124,8 +153,9 @@ class NotebookLMTool:
             Each item: {"text": string, "box_2d": [ymin, xmin, ymax, xmax] (0-1000), "font_size": int, "color": hex, "is_bold": bool}
             """
             
-            resp_layout = self.client.models.generate_content(
-                model="gemini-2.5-flash", 
+            # 使用帶重試機制的呼叫
+            resp_layout = self._call_gemini_with_retry(
+                model_name="gemini-2.0-flash", 
                 contents=[layout_prompt, img],
                 config=types.GenerateContentConfig(response_mime_type="application/json")
             )
@@ -137,10 +167,6 @@ class NotebookLMTool:
             blocks = self._extract_json(resp_layout.text)
             result["blocks"] = blocks
             
-            # 紀錄 Log
-            for b in blocks:
-                if b.get("text"): result["log"] += f"[P{page_index+1}] {b['text'][:20]}...\n"
-
         except Exception as e:
             print(f"❌ [Page {page_index+1}] Layout Error: {e}", flush=True)
             result["log"] += f"[P{page_index+1}] Layout Analysis Failed.\n"
@@ -174,54 +200,49 @@ class NotebookLMTool:
         # 2. PDF 轉圖片 (降低 DPI 加速)
         progress(0.1, desc="正在將 PDF 轉為圖片 (DPI=150)...")
         try:
-            # dpi=150 足夠螢幕檢視，且大幅減少上傳時間
             images = convert_from_path(pdf_file, dpi=150) 
         except Exception as e:
             raise ValueError(f"PDF 轉換失敗: {str(e)}")
 
-        # 3. 平行處理 (Parallel Execution)
-        # 根據 CPU 核心數或 API 限制設定 workers，建議 3-5 避免 Rate Limit
-        max_workers = 4 
-        results_map = {} # 用來存結果，確保順序正確
+        # 3. 平行處理 (Parallel Execution with Safety)
+        # 降低併發數以適應免費版 API
+        max_workers = 2 
+        results_map = {} 
 
-        progress(0.2, desc="🚀 AI 多工處理中 (可能需要稍等)...")
+        progress(0.2, desc="🚀 AI 處理中 (已啟用速率保護)...")
         
         with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
-            # 提交所有任務
-            future_to_page = {
-                executor.submit(self.process_single_page, i, img, img_output_dir): i 
-                for i, img in enumerate(images)
-            }
+            # 提交任務，但加入微小延遲避免瞬間併發過高
+            future_to_page = {}
+            for i, img in enumerate(images):
+                time.sleep(1) # 讓請求稍微錯開
+                future = executor.submit(self.process_single_page, i, img, img_output_dir)
+                future_to_page[future] = i
             
             # 等待完成
             for future in concurrent.futures.as_completed(future_to_page):
                 try:
                     res = future.result()
                     results_map[res["index"]] = res
-                    # 更新 Token
                     total_input_tokens += res["tokens_in"]
                     total_output_tokens += res["tokens_out"]
                 except Exception as exc:
                     print(f"Page processing generated an exception: {exc}")
 
-        # 4. 依序組裝 PPTX (確保順序正確)
+        # 4. 依序組裝 PPTX
         progress(0.8, desc="正在組裝 PPTX...")
         
-        cleaned_images_paths = [] # 用於 ZIP
+        cleaned_images_paths = [] 
         
         for i in range(len(images)):
             if i not in results_map:
-                print(f"Missing result for page {i}")
                 continue
-                
             res = results_map[i]
             
-            # 更新 Log 與 Preview
             full_text_log += res["log"]
             if res["preview"]: gallery_preview.append(res["preview"])
             if res["bg_path"]: cleaned_images_paths.append(res["bg_path"])
 
-            # 建立 Slide
             slide = prs.slides.add_slide(prs.slide_layouts[6])
             
             # A. 貼背景
@@ -235,7 +256,6 @@ class NotebookLMTool:
                 text_content = block.get("text", "")
                 if not text_content: continue
                 
-                # 座標轉換
                 box = block.get("box_2d", [0, 0, 100, 100])
                 ymin, xmin, ymax, xmax = box
                 left = Inches((xmin / 1000) * 16)
@@ -254,7 +274,6 @@ class NotebookLMTool:
                 p.font.bold = block.get("is_bold", False)
                 try:
                     hex_c = block.get("color", "#000000").replace("#", "")
-                    # 如果背景去字失敗，原圖背景可能很複雜，文字顏色可能需要調整 (這裡暫不處理，保持原色)
                     p.font.color.rgb = RGBColor.from_string(hex_c)
                 except: pass