Update app.py

app.py

@@ -1,17 +1,18 @@
 import gradio as gr
-import google.generativeai as genai
-import os
 import json
+import os
 import pandas as pd
 import tempfile
-from pdf2image import convert_from_path
-from pptx import Presentation
-from pptx.util import Inches, Pt
-from pptx.dml.color import RGBColor
-from huggingface_hub import HfApi, hf_hub_download
+import zipfile
+import shutil
 from dotenv import load_dotenv
+from huggingface_hub import HfApi, hf_hub_download
+from pdf2image import convert_from_path
+import google.generativeai as genai
+from google.genai import types # 確保相容性
+from PIL import Image
 
-# --- 設定與常數 ---
+# Load Env
 load_dotenv()
 PROF_SAVE_FILE = "saved_professors.json"
 COMP_SAVE_FILE = "saved_companies.json"
@@ -19,296 +20,602 @@ HF_TOKEN = os.getenv("HF_TOKEN")
 DATASET_REPO_ID = os.getenv("DATASET_REPO_ID")
 
 # ==========================================
-# 🧠 核心服務層 (The Logic / Chef)
+# 🧠 Unified AI Service (整合後端邏輯)
 # ==========================================
 class UnifiedService:
-    def __init__(self, api_key_input=None):
-        self.api_key = self._get_api_key(api_key_input)
+    def __init__(self):
+        self.api_key = self._get_api_key()
         if self.api_key:
             genai.configure(api_key=self.api_key)
-            # 使用支援 Google Search 的模型
-            self.model_name = "gemini-2.0-flash-exp" 
-
-    def _get_api_key(self, user_key):
-        if user_key and user_key.strip(): return user_key.strip()
-        system_key = os.getenv("GEMINI_API_KEY")
-        if system_key: return system_key
-        return None # 允許初始化時無 Key，但在使用功能時會噴錯
-
-    def _check_key(self):
-        if not self.api_key: raise ValueError("請先輸入 API Key 或設定系統環境變數")
-
-    # --- 1. PDF 轉 PPTX ---
-    def analyze_pdf_to_pptx(self, pdf_file, progress):
-        self._check_key()
-        model = genai.GenerativeModel(self.model_name)
-        prs = Presentation()
-        prs.slide_width = Inches(16); prs.slide_height = Inches(9)
-        
-        progress(0.1, desc="轉檔中...")
-        images = convert_from_path(pdf_file)
+            self.model_id = "gemini-2.0-flash-exp" # 使用較新的模型
+        else:
+            print("⚠️ Warning: No API Key found.")
+
+    def _get_api_key(self):
+        # 優先讀取環境變數 (Secrets)
+        return os.getenv("GEMINI_API_KEY")
+
+    def set_user_key(self, key):
+        """允許使用者在介面上暫時替換 Key"""
+        if key and key.strip():
+            self.api_key = key.strip()
+            genai.configure(api_key=self.api_key)
+
+    def _check_client(self):
+        if not self.api_key:
+            raise ValueError("API Key 未設定，請檢查 .env, Secrets 或在介面上輸入")
+
+    # --- 🛠️ New Feature: PDF 智能拆解 (NotebookLM 專用) ---
+    def decompose_pdf(self, pdf_file, progress=gr.Progress()):
+        self._check_client()
+        if not pdf_file: return None, None, "請上傳 PDF"
+
+        # 1. PDF 轉圖片
+        progress(0.1, desc="正在將 PDF 轉為圖片...")
+        try:
+            images = convert_from_path(pdf_file)
+        except Exception as e:
+            return None, None, f"PDF 轉換失敗 (請確認系統已安裝 poppler): {e}"
+
+        # 準備暫存資料夾
+        tmp_dir = tempfile.mkdtemp()
+        clean_img_dir = os.path.join(tmp_dir, "cleaned_images")
+        os.makedirs(clean_img_dir, exist_ok=True)
         
+        full_text_content = ""
+        processed_images = []
+        model = genai.GenerativeModel(self.model_id)
+
+        # 2. 逐頁處理
         for i, img in enumerate(images):
-            progress(0.1 + (0.8 * (i / len(images))), desc=f"分析第 {i+1} 頁...")
-            slide = prs.slides.add_slide(prs.slide_layouts[6])
+            progress(0.1 + (0.8 * (i / len(images))), desc=f"AI 正在拆解第 {i+1}/{len(images)} 頁...")
             
-            prompt = "Detect all text blocks. Return JSON: [{'text':..., 'box_2d':[ymin,xmin,ymax,xmax] (0-1000), 'font_size':int, 'is_bold':bool, 'color':hex}]"
+            # Action A: 提取文字 (OCR)
             try:
-                response = model.generate_content([prompt, img], generation_config={"response_mime_type": "application/json"})
-                blocks = json.loads(response.text)
-                for b in blocks:
-                    box = b.get("box_2d", [0,0,0,0])
-                    left, top = Inches((box[1]/1000)*16), Inches((box[0]/1000)*9)
-                    width, height = Inches(((box[3]-box[1])/1000)*16), Inches(((box[2]-box[0])/1000)*9)
-                    tx = slide.shapes.add_textbox(left, top, width, height)
-                    p = tx.text_frame.paragraphs[0]
-                    p.text = b.get("text",""); p.font.size = Pt(b.get("font_size", 12)); p.font.bold = b.get("is_bold", False)
-                    try: p.font.color.rgb = RGBColor.from_string(b.get("color", "#000000").replace("#",""))
-                    except: pass
-            except Exception as e: print(f"Page {i} err: {e}")
+                prompt_ocr = "Extract all text content from this image strictly. Do not describe the layout."
+                ocr_resp = model.generate_content([prompt_ocr, img])
+                page_text = ocr_resp.text
+            except:
+                page_text = "[Text Extraction Failed]"
             
-        out = tempfile.mktemp(suffix=".pptx")
-        prs.save(out)
-        return out, "✅ 轉換完成"
-
-    # --- 2. 圖片去字 ---
-    def remove_text(self, image):
-        self._check_key()
-        model = genai.GenerativeModel(self.model_name)
-        prompt = "Remove all text from this image, fill background naturally. Return image only."
-        resp = model.generate_content([prompt, image]) # V1 SDK 通常回傳 multipart，這裡簡化處理
-        # 注意: Gemini V1 SDK 在 Python 直接回傳 image 比較 tricky，若失敗建議檢查 SDK 版本
-        # 這裡假設環境支援直接回圖，若否則需用 requests 操作 REST API
-        try:
-            return resp.parts[0].image 
-        except:
-            return image # Fallback
-
-    # --- 3. 搜尋 (教授/公司) 共用邏輯 ---
-    def _search_with_google(self, query, prompt_template):
-        self._check_key()
-        # 這裡使用 Google Search Tool 設定
+            full_text_content += f"--- Page {i+1} ---\n{page_text}\n\n"
+
+            # Action B: 移除文字 (In-painting)
+            try:
+                prompt_clean = "Remove all text from this image and fill in the background naturally. Return only the image."
+                clean_resp = model.generate_content([prompt_clean, img])
+                # 嘗試取得圖片 (處理 V1/V2 SDK 差異)
+                try:
+                    clean_img = clean_resp.parts[0].image
+                except:
+                    # Fallback 若 SDK 版本不同或回傳格式不同
+                    clean_img = img # 若失敗則保留原圖
+                
+                # 存檔
+                img_filename = f"page_{i+1:03d}_clean.png"
+                img_path = os.path.join(clean_img_dir, img_filename)
+                clean_img.save(img_path)
+                processed_images.append(clean_img)
+            except Exception as e:
+                print(f"Clean Error on page {i}: {e}")
+                processed_images.append(img)
+
+        # 3. 打包結果
+        progress(0.9, desc="正在打包檔案...")
+        
+        # 儲存文字檔
+        txt_path = os.path.join(tmp_dir, "extracted_text.txt")
+        with open(txt_path, "w", encoding="utf-8") as f:
+            f.write(full_text_content)
+
+        # 建立 ZIP
+        zip_path = os.path.join(tmp_dir, "notebooklm_result.zip")
+        with zipfile.ZipFile(zip_path, 'w', zipfile.ZIP_DEFLATED) as zf:
+            zf.write(txt_path, "content.txt")
+            for root, dirs, files in os.walk(clean_img_dir):
+                for file in files:
+                    zf.write(os.path.join(root, file), os.path.join("images", file))
+
+        return zip_path, full_text_content, processed_images
+
+    # --- 🎓 Professor Search Logic (Copied from original) ---
+    def search_professors(self, query, exclude_names=[]):
+        self._check_client()
+        exclusion = f"IMPORTANT: Do not include: {', '.join(exclude_names)}." if exclude_names else ""
+        
+        # Phase 1: Search
         tools = [{"google_search": {}}]
-        model = genai.GenerativeModel(self.model_name, tools=tools)
+        model_tools = genai.GenerativeModel(self.model_id, tools=tools)
         
-        # Step 1: Search
-        resp1 = model.generate_content(prompt_template.format(query=query))
+        prompt = f"""
+        Using Google Search, find 10 prominent professors in universities across Taiwan who are experts in "{query}".
+        FACT CHECK: Must be current faculty. {exclusion}
+        List them (Name - University - Department) in Traditional Chinese.
+        """
+        resp1 = model_tools.generate_content(prompt)
         
-        # Step 2: Extract JSON (Pure Text Model)
-        model_extract = genai.GenerativeModel(self.model_name) # No tools for extraction
-        extract_prompt = f"Extract structured data from this text into JSON array: {resp1.text}"
-        resp2 = model_extract.generate_content(extract_prompt, generation_config={"response_mime_type": "application/json"})
+        # Phase 2: Extract JSON
+        model_pure = genai.GenerativeModel(self.model_id)
+        extract_prompt = f"""
+        Extract professor names, universities, and departments from the text below.
+        Return ONLY a JSON array: [{{"name": "...", "university": "...", "department": "...", "tags": ["tag1"]}}]
+        Text: {resp1.text}
+        """
+        resp2 = model_pure.generate_content(extract_prompt, generation_config={"response_mime_type": "application/json"})
         try: return json.loads(resp2.text)
         except: return []
 
-    def search_professors(self, query):
-        p = "Find 10 prominent professors in Taiwan for '{query}'. Return name, university, department."
-        return self._search_with_google(query, p)
+    def get_professor_details(self, professor):
+        self._check_client()
+        tools = [{"google_search": {}}]
+        model = genai.GenerativeModel(self.model_id, tools=tools)
+        prompt = f"Act as academic consultant. Investigate Professor {professor.get('name')} from {professor.get('university')}. Find key publications and industry projects. Report in Traditional Chinese Markdown."
+        resp = model.generate_content(prompt)
+        return self._format_response_with_sources(resp)
 
-    def search_companies(self, query):
-        p = "Find 5-10 Taiwanese companies for '{query}'. Return name, industry."
-        return self._search_with_google(query, p)
+    # --- 🏢 Company Search Logic (Copied from original) ---
+    def search_companies(self, query, exclude_names=[]):
+        self._check_client()
+        exclusion = f"IMPORTANT: Do not include: {', '.join(exclude_names)}." if exclude_names else ""
+        
+        tools = [{"google_search": {}}]
+        model = genai.GenerativeModel(self.model_id, tools=tools)
+        prompt = f"""
+        Using Google Search, find 5-10 Taiwanese companies related to: "{query}".
+        {exclusion}
+        List them (Name - Industry) in Traditional Chinese.
+        """
+        resp1 = model.generate_content(prompt)
+        
+        model_pure = genai.GenerativeModel(self.model_id)
+        extract_prompt = f"""
+        Extract company names and industry from text.
+        Return ONLY JSON array: [{{"name": "...", "industry": "...", "tags": ["tag1"]}}]
+        Text: {resp1.text}
+        """
+        resp2 = model_pure.generate_content(extract_prompt, generation_config={"response_mime_type": "application/json"})
+        try: return json.loads(resp2.text)
+        except: return []
 
-    def get_details(self, data, role):
-        self._check_key()
+    def get_company_details(self, company):
+        self._check_client()
         tools = [{"google_search": {}}]
-        model = genai.GenerativeModel(self.model_name, tools=tools)
-        prompt = f"Act as {role}. Investigate: {json.dumps(data)}. Report in Traditional Chinese Markdown."
+        model = genai.GenerativeModel(self.model_id, tools=tools)
+        prompt = f"Act as Business Analyst. Investigate company: '{company.get('name')}'. Focus on products, culture, and disputes. Report in Traditional Chinese Markdown."
         resp = model.generate_content(prompt)
+        return self._format_response_with_sources(resp)
+
+    # --- Shared Helpers ---
+    def chat_with_ai(self, history, msg, context, role):
+        self._check_client()
+        model = genai.GenerativeModel(self.model_id)
+        sys_prompt = f"{role}:\nContext: {context}"
         
-        # 處理來源引用 (V1 SDK)
-        sources = []
-        if hasattr(resp.candidates[0], 'grounding_metadata'):
-            chunks = resp.candidates[0].grounding_metadata.grounding_chunks
-            for c in chunks:
-                if c.web: sources.append({"title": c.web.title, "uri": c.web.uri})
-        
-        # 去重
-        unique_sources = list({v['uri']:v for v in sources}.values())
-        return {"text": resp.text, "sources": unique_sources}
-
-    def chat(self, hist, msg, context, role):
-        self._check_key()
-        model = genai.GenerativeModel(self.model_name)
-        chat = model.start_chat(history=[
-            {"role": "user" if h[0] else "model", "parts": [h[0] or h[1]]} for h in hist
-        ])
-        full_msg = f"Context: {context}\nInstruction: {role}\nUser: {msg}"
-        resp = chat.send_message(full_msg)
+        # Convert history for Gemini
+        chat_hist = []
+        for h in history:
+            chat_hist.append({"role": "user", "parts": [h[0]]})
+            if len(h) > 1: chat_hist.append({"role": "model", "parts": [h[1]]})
+            
+        chat = model.start_chat(history=chat_hist)
+        resp = chat.send_message(f"{sys_prompt}\nUser: {msg}")
         return resp.text
 
-# ==========================================
-# 💾 資料存取層 (Persistence)
-# ==========================================
+    def _format_response_with_sources(self, response):
+        sources = []
+        if hasattr(response.candidates[0], 'grounding_metadata'):
+            gm = response.candidates[0].grounding_metadata
+            if hasattr(gm, 'grounding_chunks'):
+                for chunk in gm.grounding_chunks:
+                    if hasattr(chunk, 'web'):
+                        sources.append({"title": chunk.web.title, "uri": chunk.web.uri})
+        # Deduplicate
+        unique_sources = list({v['uri']: v for v in sources}.values())
+        return {"text": response.text, "sources": unique_sources}
+
+# Init Service
+gemini_service = UnifiedService()
+
+# --- Helper Functions (Preserved from your code) ---
+
 def load_data(filename):
+    data = []
     if HF_TOKEN and DATASET_REPO_ID:
         try: hf_hub_download(repo_id=DATASET_REPO_ID, filename=filename, repo_type="dataset", token=HF_TOKEN, local_dir=".")
         except: pass
     if os.path.exists(filename):
         try: 
-            with open(filename, 'r', encoding='utf-8') as f: return json.load(f)
-        except: pass
-    return []
+            with open(filename, 'r', encoding='utf-8') as f: data = json.load(f)
+        except: data = []
+    return data
 
 def save_data(data, filename):
-    with open(filename, 'w', encoding='utf-8') as f: json.dump(data, f, ensure_ascii=False, indent=2)
+    try:
+        with open(filename, 'w', encoding='utf-8') as f: json.dump(data, f, ensure_ascii=False, indent=2)
+    except: return
     if HF_TOKEN and DATASET_REPO_ID:
         try:
             api = HfApi(token=HF_TOKEN)
-            api.upload_file(path_or_fileobj=filename, path_in_repo=filename, repo_id=DATASET_REPO_ID, repo_type="dataset", commit_message="Sync")
+            api.upload_file(path_or_fileobj=filename, path_in_repo=filename, repo_id=DATASET_REPO_ID, repo_type="dataset", commit_message=f"Sync {filename}")
         except: pass
 
-# ==========================================
-# 🖥️ 介面邏輯 (UI Helpers)
-# ==========================================
-def format_df(data_list, cols):
-    if not data_list: return pd.DataFrame(columns=cols)
-    res = []
-    for d in data_list:
-        icon = {'match':'✅','good':'✅','risk':'⚠️'}.get(d.get('status'),'')
-        res.append([f"{icon} {d.get('name')}", d.get('university') or d.get('industry'), ", ".join(d.get('tags',[]))])
-    return pd.DataFrame(res, columns=cols)
+def get_tags_text(item):
+    if not item or not item.get('tags'): return "目前標籤: (無)"
+    return "🏷️ " + ", ".join([f"`{t}`" for t in item['tags']])
 
-# ==========================================
-# 🚀 主程式 (Gradio)
-# ==========================================
-def main_app():
-    # 初始化
-    prof_data = load_data(PROF_SAVE_FILE)
-    comp_data = load_data(COMP_SAVE_FILE)
+def get_tags_choices(item): return item.get('tags', []) if item else []
+def prof_get_key(p): return f"{p['name']}-{p['university']}"
+def comp_get_key(c): return f"{c['name']}"
+
+def prof_format_df(source_list, saved_list):
+    if not source_list: return pd.DataFrame(columns=["狀態", "姓名", "大學", "系所", "標籤"])
+    if saved_list is None: saved_list = []
+    saved_map = {prof_get_key(p): p for p in saved_list}
+    data = []
+    for p in source_list:
+        dp = saved_map.get(prof_get_key(p), p)
+        icon = {'match':'✅','mismatch':'❌','pending':'❓'}.get(dp.get('status'), '')
+        detail = "📄" if dp.get('details') else ""
+        data.append([f"{icon} {detail}", dp['name'], dp['university'], dp['department'], ", ".join(dp.get('tags', []))])
+    return pd.DataFrame(data, columns=["狀態", "姓名", "大學", "系所", "標籤"])
+
+def comp_format_df(source_list, saved_list):
+    if not source_list: return pd.DataFrame(columns=["狀態", "公司名稱", "產業類別", "標籤"])
+    if saved_list is None: saved_list = []
+    saved_map = {comp_get_key(c): c for c in saved_list}
+    data = []
+    for c in source_list:
+        dc = saved_map.get(comp_get_key(c), c)
+        icon = {'good':'✅','risk':'⚠️','pending':'❓'}.get(dc.get('status'), '')
+        detail = "📄" if dc.get('details') else ""
+        data.append([f"{icon} {detail}", dc['name'], dc.get('industry','未知'), ", ".join(dc.get('tags', []))])
+    return pd.DataFrame(data, columns=["狀態", "公司名稱", "產業類別", "標籤"])
+
+# --- Wrappers for Prof Logic ---
+def prof_search(query, current_saved):
+    if not query: return gr.update(), current_saved, gr.update()
+    try:
+        res = gemini_service.search_professors(query)
+        return prof_format_df(res, current_saved), res, gr.update(visible=True)
+    except Exception as e: raise gr.Error(f"搜尋失敗: {e}")
+
+def prof_load_more(query, cur_res, cur_saved):
+    if not query: return gr.update(), cur_res
+    try:
+        new_res = gemini_service.search_professors(query, exclude_names=[p['name'] for p in cur_res])
+        exist_keys = set(prof_get_key(p) for p in cur_res)
+        for p in new_res:
+            if prof_get_key(p) not in exist_keys: cur_res.append(p)
+        return prof_format_df(cur_res, cur_saved), cur_res
+    except Exception as e: raise gr.Error(f"載入失敗: {e}")
+
+def prof_select(evt: gr.SelectData, search_res, saved_data, view_mode):
+    if not evt: return [gr.update()]*8
+    idx = evt.index[0]
+    target = saved_data if view_mode == "追蹤清單" else search_res
+    if not target or idx >= len(target): return [gr.update()]*8
+    p = target[idx]
+    key = prof_get_key(p)
+    saved_p = next((x for x in saved_data if prof_get_key(x) == key), None)
+    curr = saved_p if saved_p else p
+    md = ""
+    if curr.get('details') and len(curr.get('details')) > 10:
+        md = curr['details']
+        if not saved_p: saved_data.insert(0, curr); save_data(saved_data, PROF_SAVE_FILE)
+    else:
+        gr.Info(f"正在調查 {curr['name']}...")
+        try:
+            res = gemini_service.get_professor_details(curr)
+            curr['details'] = res['text']; curr['sources'] = res['sources']
+            md = res['text']
+            if saved_p: saved_p.update(curr)
+            else: saved_data.insert(0, curr)
+            save_data(saved_data, PROF_SAVE_FILE)
+        except Exception as e: raise gr.Error(f"調查失敗: {e}")
+    if curr.get('sources'): md += "\n\n### 📚 參考來源\n" + "\n".join([f"- [{s['title']}]({s['uri']})" for s in curr['sources']])
+    return gr.update(visible=True), md, [], curr, saved_data, get_tags_text(curr), gr.update(choices=get_tags_choices(curr), value=None), gr.update(visible=True)
+
+def prof_chat(hist, msg, curr):
+    if not curr: return hist, ""
+    try:
+        reply = gemini_service.chat_with_ai(hist, msg, curr.get('details', ''), "你是學術顧問，請根據這份教授資料回答")
+        hist.append((msg, reply))
+    except Exception as e: hist.append((msg, f"Error: {e}"))
+    return hist, ""
+
+def prof_add_tag(tag, curr, saved, mode, res):
+    if not curr or not tag: return gr.update(), gr.update(), gr.update(), saved, gr.update()
+    if 'tags' not in curr: curr['tags'] = []
+    if tag not in curr['tags']:
+        curr['tags'].append(tag)
+        key = prof_get_key(curr)
+        found = False
+        for i, p in enumerate(saved):
+            if prof_get_key(p) == key: saved[i] = curr; found=True; break
+        if not found: saved.insert(0, curr)
+        save_data(saved, PROF_SAVE_FILE)
+    return gr.update(value=""), get_tags_text(curr), gr.update(choices=curr['tags']), saved, prof_format_df(saved if mode=="追蹤清單" else res, saved)
+
+def prof_remove_tag(tag, curr, saved, mode, res):
+    if not curr or not tag: return gr.update(), gr.update(), saved, gr.update()
+    if 'tags' in curr and tag in curr['tags']:
+        curr['tags'].remove(tag)
+        key = prof_get_key(curr)
+        for i, p in enumerate(saved):
+            if prof_get_key(p) == key: saved[i] = curr; break
+        save_data(saved, PROF_SAVE_FILE)
+    return get_tags_text(curr), gr.update(choices=curr['tags'], value=None), saved, prof_format_df(saved if mode=="追蹤清單" else res, saved)
+
+def prof_update_status(stat, curr, saved, mode, res):
+    if not curr: return gr.update(), saved
+    curr['status'] = stat if curr.get('status') != stat else None
+    key = prof_get_key(curr)
+    for i, p in enumerate(saved):
+        if prof_get_key(p) == key: saved[i] = curr; break
+    save_data(saved, PROF_SAVE_FILE)
+    return prof_format_df(saved if mode=="追蹤清單" else res, saved), saved
+
+def prof_remove(curr, saved, mode, res):
+    if not curr: return gr.update(), gr.update(value=None), saved, gr.update(visible=False)
+    key = prof_get_key(curr)
+    new_saved = [p for p in saved if prof_get_key(p) != key]
+    save_data(new_saved, PROF_SAVE_FILE)
+    return gr.Info("已移除"), prof_format_df(new_saved if mode=="追蹤清單" else res, new_saved), new_saved, gr.update(visible=False)
+
+def prof_toggle(mode, res, saved):
+    return prof_format_df(res if mode=="搜尋結果" else saved, saved), gr.update(visible=mode=="搜尋結果")
+
+# --- Wrappers for Company Logic ---
+def comp_search(query, current_saved):
+    if not query: return gr.update(), current_saved, gr.update()
+    try:
+        res = gemini_service.search_companies(query)
+        return comp_format_df(res, current_saved), res, gr.update(visible=True)
+    except Exception as e: raise gr.Error(f"搜尋失敗: {e}")
+
+def comp_load_more(query, cur_res, cur_saved):
+    if not query: return gr.update(), cur_res
+    try:
+        new_res = gemini_service.search_companies(query, exclude_names=[c['name'] for c in cur_res])
+        exist_keys = set(comp_get_key(c) for c in cur_res)
+        for c in new_res:
+            if comp_get_key(c) not in exist_keys: cur_res.append(c)
+        return comp_format_df(cur_res, cur_saved), cur_res
+    except Exception as e: raise gr.Error(f"載入失敗: {e}")
+
+def comp_select(evt: gr.SelectData, search_res, saved_data, view_mode):
+    if not evt: return [gr.update()]*8
+    idx = evt.index[0]
+    target = saved_data if view_mode == "追蹤清單" else search_res
+    if not target or idx >= len(target): return [gr.update()]*8
+    c = target[idx]
+    key = comp_get_key(c)
+    saved_c = next((x for x in saved_data if comp_get_key(x) == key), None)
+    curr = saved_c if saved_c else c
+    md = ""
+    if curr.get('details') and len(curr.get('details')) > 10:
+        md = curr['details']
+        if not saved_c: saved_data.insert(0, curr); save_data(saved_data, COMP_SAVE_FILE)
+    else:
+        gr.Info(f"正在調查 {curr['name']}...")
+        try:
+            res = gemini_service.get_company_details(curr)
+            curr['details'] = res['text']; curr['sources'] = res['sources']
+            md = res['text']
+            if saved_c: saved_c.update(curr)
+            else: saved_data.insert(0, curr)
+            save_data(saved_data, COMP_SAVE_FILE)
+        except Exception as e: raise gr.Error(f"調查失敗: {e}")
+    if curr.get('sources'): md += "\n\n### 📚 資料來源\n" + "\n".join([f"- [{s['title']}]({s['uri']})" for s in curr['sources']])
+    return gr.update(visible=True), md, [], curr, saved_data, get_tags_text(curr), gr.update(choices=get_tags_choices(curr), value=None), gr.update(visible=True)
+
+def comp_chat(hist, msg, curr):
+    if not curr: return hist, ""
+    try:
+        reply = gemini_service.chat_with_ai(hist, msg, curr.get('details', ''), "你是商業顧問，請根據這份公司調查報告回答")
+        hist.append((msg, reply))
+    except Exception as e: hist.append((msg, f"Error: {e}"))
+    return hist, ""
+
+def comp_add_tag(tag, curr, saved, mode, res):
+    if not curr or not tag: return gr.update(), gr.update(), gr.update(), saved, gr.update()
+    if 'tags' not in curr: curr['tags'] = []
+    if tag not in curr['tags']:
+        curr['tags'].append(tag)
+        key = comp_get_key(curr)
+        found = False
+        for i, c in enumerate(saved):
+            if comp_get_key(c) == key: saved[i] = curr; found=True; break
+        if not found: saved.insert(0, curr)
+        save_data(saved, COMP_SAVE_FILE)
+    return gr.update(value=""), get_tags_text(curr), gr.update(choices=curr['tags']), saved, comp_format_df(saved if mode=="追蹤清單" else res, saved)
+
+def comp_remove_tag(tag, curr, saved, mode, res):
+    if not curr or not tag: return gr.update(), gr.update(), saved, gr.update()
+    if 'tags' in curr and tag in curr['tags']:
+        curr['tags'].remove(tag)
+        key = comp_get_key(curr)
+        for i, c in enumerate(saved):
+            if comp_get_key(c) == key: saved[i] = curr; break
+        save_data(saved, COMP_SAVE_FILE)
+    return get_tags_text(curr), gr.update(choices=curr['tags'], value=None), saved, comp_format_df(saved if mode=="追蹤清單" else res, saved)
+
+def comp_update_status(stat, curr, saved, mode, res):
+    if not curr: return gr.update(), saved
+    curr['status'] = stat if curr.get('status') != stat else None
+    key = comp_get_key(curr)
+    for i, c in enumerate(saved):
+        if comp_get_key(c) == key: saved[i] = curr; break
+    save_data(saved, COMP_SAVE_FILE)
+    return comp_format_df(saved if mode=="追蹤清單" else res, saved), saved
+
+def comp_remove(curr, saved, mode, res):
+    if not curr: return gr.update(), gr.update(value=None), saved, gr.update(visible=False)
+    key = comp_get_key(curr)
+    new_saved = [c for c in saved if comp_get_key(c) != key]
+    save_data(new_saved, COMP_SAVE_FILE)
+    return gr.Info("已移除"), comp_format_df(new_saved if mode=="追蹤清單" else res, new_saved), new_saved, gr.update(visible=False)
 
-    with gr.Blocks(title="Prof.404 x PPT.404 Ultimate", theme=gr.themes.Soft()) as demo:
+def comp_toggle(mode, res, saved):
+    return comp_format_df(res if mode=="搜尋結果" else saved, saved), gr.update(visible=mode=="搜尋結果")
+
+# Init
+def prof_init(): d = load_data(PROF_SAVE_FILE); return d, prof_format_df(d, d)
+def comp_init(): d = load_data(COMP_SAVE_FILE); return d, comp_format_df(d, d)
+
+
+# ==========================
+# 🖥️ UI Layout (Modified)
+# ==========================
+with gr.Blocks(title="Prof.404.Com 產學導航系統", theme=gr.themes.Soft()) as demo:
+    
+    gr.Markdown("""
+    <div align="center">
+    
+    # 🚀 Prof.404.Com 產學導航系統 (含 NotebookLM 擴充工具)
+    **學術研究啟程、產業導航、以及您的文件處理瑞士刀**
+    </div>
+    """)
+    
+    with gr.Accordion("🔑 API Key 設定", open=False):
+        api_input = gr.Textbox(label="Gemini API Key", placeholder="若未設定環境變數，請在此輸入", type="password")
+        api_btn = gr.Button("設定 Key")
+        api_btn.click(lambda k: gemini_service.set_user_key(k), inputs=api_input)
+
+    with gr.Tabs():
         
-        # 全域 Key
-        with gr.Accordion("🔑 系統設定 (API Key)", open=False):
-            api_key = gr.Textbox(label="Google Gemini API Key", type="password", placeholder="若未填寫則使用系統預設")
-
-        gr.Markdown(
-            """
-            <div align="center">
-            <h1>🚀 Prof.404 Ultimate: 產學導航 & 文件工具站</h1>
-            <h3>整合文件視覺處理 (PPT/Img) 與 產學資源導航 (Prof/Com) 的全方位平台</h3>
-            </div>
-            """
-        )
-
-        with gr.Tabs():
+        # ==========================
+        # Tab 1: 🛠️ 工具箱 (PDF 智能拆解)
+        # ==========================
+        with gr.Tab("🛠️ NotebookLM 拆解工具"):
+            gr.Markdown("### 📄 PDF 智能拆解 (文字/圖片分離)")
+            gr.Markdown("上傳 NotebookLM 生成的 PDF，AI 將自動為您：**1. 提取全文文字** | **2. 移除圖片中的文字(還原背景)**")
             
-            # --- Tab 1: 工具箱 ---
-            with gr.Tab("🛠️ 文件工具箱 (PPT.404)"):
-                with gr.Row():
-                    with gr.Column():
-                        gr.Markdown("### 📄 PDF 轉 PPTX (含排版還原)")
-                        pdf_file = gr.File(label="上傳 PDF")
-                        pdf_btn = gr.Button("開始轉換", variant="primary")
-                        ppt_out = gr.File(label="下載 PPTX")
-                        pdf_msg = gr.Textbox(label="狀態", interactive=False)
-                        
-                        pdf_btn.click(
-                            lambda f, k: UnifiedService(k).analyze_pdf_to_pptx(f, gr.Progress()),
-                            inputs=[pdf_file, api_key], outputs=[ppt_out, pdf_msg]
-                        )
-
-                    with gr.Column():
-                        gr.Markdown("### 🎨 圖片智慧去字")
-                        img_in = gr.Image(type="pil", label="原圖")
-                        img_btn = gr.Button("一鍵去除", variant="primary")
-                        img_out = gr.Image(label="結果")
-                        
-                        img_btn.click(
-                            lambda i, k: UnifiedService(k).remove_text(i),
-                            inputs=[img_in, api_key], outputs=[img_out]
-                        )
-
-            # --- Tab 2: 找教授 ---
-            with gr.Tab("🎓 找教授 (Prof.404)"):
-                p_state = gr.State(prof_data)
-                p_current = gr.State(None) # 當前選中的教授
-                
-                with gr.Row():
-                    p_query = gr.Textbox(label="搜尋領域", scale=4)
-                    p_btn = gr.Button("搜尋", scale=1)
-                
-                with gr.Row():
-                    p_table = gr.Dataframe(headers=["姓名", "大學", "標籤"], interactive=False, scale=1)
-                    with gr.Column(scale=1, visible=False) as p_detail_col:
-                        p_md = gr.Markdown()
-                        p_chat = gr.Chatbot(height=300)
-                        p_msg = gr.Textbox(label="詢問關於此教授")
+            with gr.Row():
+                with gr.Column(scale=1):
+                    pdf_input = gr.File(label="上傳 PDF (來自 NotebookLM 或其他)")
+                    process_btn = gr.Button("🚀 開始一鍵拆解", variant="primary")
                 
-                # Logic Wrappers
-                def search_p(q, k, saved):
-                    svc = UnifiedService(k)
-                    res = svc.search_professors(q)
-                    return res, format_df(res, ["姓名","大學","標籤"])
-
-                def select_p(evt: gr.SelectData, res, k, saved):
-                    svc = UnifiedService(k)
-                    item = res[evt.index[0]]
-                    # 取得詳細資料
-                    det = svc.get_details(item, "Academic Consultant")
-                    item['details'] = det['text']
-                    # 簡易儲存邏輯 (為了Demo簡化，實際建議加上去重)
-                    saved.append(item) 
-                    save_data(saved, PROF_SAVE_FILE)
-                    
-                    display_text = det['text'] + "\n\n📚 來源:\n" + "\n".join([f"- {s['title']}" for s in det['sources']])
-                    return gr.update(visible=True), display_text, [], item, saved
-
-                def chat_p(hist, msg, item, k):
-                    svc = UnifiedService(k)
-                    reply = svc.chat(hist, msg, item.get('details'), "Academic Consultant")
-                    hist.append((msg, reply))
-                    return hist, ""
-
-                p_btn.click(search_p, [p_query, api_key, p_state], [p_state, p_table])
-                p_table.select(select_p, [p_state, api_key, p_state], [p_detail_col, p_md, p_chat, p_current, p_state])
-                p_msg.submit(chat_p, [p_chat, p_msg, p_current, api_key], [p_chat, p_msg])
-
-            # --- Tab 3: 找公司 ---
-            with gr.Tab("🏢 找公司 (Com.404)"):
-                c_state = gr.State(comp_data)
-                c_current = gr.State(None)
+                with gr.Column(scale=2):
+                    zip_output = gr.File(label="📦 下載結果 (含 clean images 與 text)")
+                    text_preview = gr.Textbox(label="📝 文字內容預覽", lines=10, max_lines=20)
+            
+            gr.Markdown("#### 🖼️ 去字後圖片預覽 (Cleaned Images)")
+            gallery_output = gr.Gallery(label="背景還原預覽", columns=4)
+
+            process_btn.click(
+                gemini_service.decompose_pdf,
+                inputs=[pdf_input],
+                outputs=[zip_output, text_preview, gallery_output]
+            )
+
+        # ==========================
+        # Tab 2: 🎓 教授去哪兒？ (保留原功能)
+        # ==========================
+        with gr.Tab("🎓 找教授 (Prof.404)"):
+            prof_saved = gr.State([])
+            prof_res = gr.State([])
+            prof_sel = gr.State(None)
+            
+            with gr.Row():
+                p_in = gr.Textbox(label="搜尋教授", placeholder="輸入研究領域...", scale=4)
+                p_btn = gr.Button("🔍 搜尋", variant="primary", scale=1)
+            
+            p_view = gr.Radio(["搜尋結果", "追蹤清單"], label="顯示模式", value="追蹤清單")
+            
+            with gr.Row():
+                with gr.Column(scale=1):
+                    p_df = gr.Dataframe(headers=["狀態","姓名","大學","系所","標籤"], datatype=["str","str","str","str","str"], interactive=False)
+                    p_load = gr.Button("載入更多", visible=False)
                 
-                with gr.Row():
-                    c_query = gr.Textbox(label="搜尋產業/公司", scale=4)
-                    c_btn = gr.Button("搜尋", scale=1)
+                with gr.Column(scale=2, visible=False) as p_col:
+                    p_md = gr.Markdown("...")
+                    with gr.Column():
+                        gr.Markdown("### 🤖 學術顧問")
+                        p_chat = gr.Chatbot(height=250)
+                        with gr.Row():
+                            p_msg = gr.Textbox(label="提問", scale=4)
+                            p_send = gr.Button("送出", scale=1)
+                    gr.Markdown("---")
+                    with gr.Column(visible=False) as p_tag_row:
+                        p_tag_disp = gr.Markdown("標籤: (無)")
+                        with gr.Row():
+                            p_tag_in = gr.Textbox(label="新增標籤", scale=3)
+                            p_tag_add = gr.Button("➕", scale=1)
+                        with gr.Accordion("刪除標籤", open=False):
+                            with gr.Row():
+                                p_tag_drop = gr.Dropdown(label="選擇標籤", choices=[], scale=3)
+                                p_tag_del = gr.Button("🗑️", scale=1, variant="secondary")
+                    with gr.Row():
+                        p_good = gr.Button("✅ 符合")
+                        p_bad = gr.Button("❌ 不符")
+                        p_pend = gr.Button("❓ 待觀察")
+                        p_rem = gr.Button("🗑️ 移除", variant="stop")
+
+            demo.load(prof_init, None, [prof_saved, p_df])
+            p_btn.click(prof_search, [p_in, prof_saved], [p_df, prof_res, p_load]).then(lambda: gr.update(value="搜尋結果"), outputs=[p_view])
+            p_load.click(prof_load_more, [p_in, prof_res, prof_saved], [p_df, prof_res])
+            p_view.change(prof_toggle, [p_view, prof_res, prof_saved], [p_df, p_load])
+            p_df.select(prof_select, [prof_res, prof_saved, p_view], [p_col, p_md, p_chat, prof_sel, prof_saved, p_tag_disp, p_tag_drop, p_tag_row])
+            p_send.click(prof_chat, [p_chat, p_msg, prof_sel], [p_chat, p_msg]); p_msg.submit(prof_chat, [p_chat, p_msg, prof_sel], [p_chat, p_msg])
+            p_tag_add.click(prof_add_tag, [p_tag_in, prof_sel, prof_saved, p_view, prof_res], [p_tag_in, p_tag_disp, p_tag_drop, prof_saved, p_df])
+            p_tag_del.click(prof_remove_tag, [p_tag_drop, prof_sel, prof_saved, p_view, prof_res], [p_tag_disp, p_tag_drop, prof_saved, p_df])
+            for btn, s in [(p_good,'match'),(p_bad,'mismatch'),(p_pend,'pending')]: btn.click(prof_update_status, [gr.State(s), prof_sel, prof_saved, p_view, prof_res], [p_df, prof_saved])
+            p_rem.click(prof_remove, [prof_sel, prof_saved, p_view, prof_res], [gr.State(None), p_df, prof_saved, p_col])
+
+        # ==========================
+        # Tab 3: 🏢 公司去那兒？ (保留原功能)
+        # ==========================
+        with gr.Tab("🏢 找公司 (Com.404)"):
+            comp_saved = gr.State([])
+            comp_res = gr.State([])
+            comp_sel = gr.State(None)
+            
+            with gr.Row():
+                c_in = gr.Textbox(label="搜尋公司/領域", placeholder="輸入產業或公司...", scale=4)
+                c_btn = gr.Button("🔍 搜尋", variant="primary", scale=1)
+            
+            c_view = gr.Radio(["搜尋結果", "追蹤清單"], label="顯示模式", value="追蹤清單")
+            
+            with gr.Row():
+                with gr.Column(scale=1):
+                    c_df = gr.Dataframe(headers=["狀態","公司名稱","產業類別","標籤"], datatype=["str","str","str","str"], interactive=False)
+                    c_load = gr.Button("載入更多", visible=False)
                 
-                with gr.Row():
-                    c_table = gr.Dataframe(headers=["公司", "產業", "標籤"], interactive=False, scale=1)
-                    with gr.Column(scale=1, visible=False) as c_detail_col:
-                        c_md = gr.Markdown()
-                        c_chat = gr.Chatbot(height=300)
-                        c_msg = gr.Textbox(label="詢問關於此公司")
-
-                # Logic Wrappers (Similar structure)
-                def search_c(q, k, saved):
-                    svc = UnifiedService(k)
-                    res = svc.search_companies(q)
-                    return res, format_df(res, ["公司","產業","標籤"])
-
-                def select_c(evt: gr.SelectData, res, k, saved):
-                    svc = UnifiedService(k)
-                    item = res[evt.index[0]]
-                    det = svc.get_details(item, "Business Analyst")
-                    item['details'] = det['text']
-                    saved.append(item)
-                    save_data(saved, COMP_SAVE_FILE)
-                    
-                    display_text = det['text'] + "\n\n📚 來源:\n" + "\n".join([f"- {s['title']}" for s in det['sources']])
-                    return gr.update(visible=True), display_text, [], item, saved
-
-                def chat_c(hist, msg, item, k):
-                    svc = UnifiedService(k)
-                    reply = svc.chat(hist, msg, item.get('details'), "Business Analyst")
-                    hist.append((msg, reply))
-                    return hist, ""
-
-                c_btn.click(search_c, [c_query, api_key, c_state], [c_state, c_table])
-                c_table.select(select_c, [c_state, api_key, c_state], [c_detail_col, c_md, c_chat, c_current, c_state])
-                c_msg.submit(chat_c, [c_chat, c_msg, c_current, api_key], [c_chat, c_msg])
-
-    demo.queue().launch()
+                with gr.Column(scale=2, visible=False) as c_col:
+                    c_md = gr.Markdown("...")
+                    with gr.Column():
+                        gr.Markdown("### 🤖 商業顧問")
+                        c_chat = gr.Chatbot(height=250)
+                        with gr.Row():
+                            c_msg = gr.Textbox(label="提問", scale=4)
+                            c_send = gr.Button("送出", scale=1)
+                    gr.Markdown("---")
+                    with gr.Column(visible=False) as c_tag_row:
+                        c_tag_disp = gr.Markdown("標籤: (無)")
+                        with gr.Row():
+                            c_tag_in = gr.Textbox(label="新增標籤", scale=3)
+                            c_tag_add = gr.Button("➕", scale=1)
+                        with gr.Accordion("刪除標籤", open=False):
+                            with gr.Row():
+                                c_tag_drop = gr.Dropdown(label="選擇標籤", choices=[], scale=3)
+                                c_tag_del = gr.Button("🗑️", scale=1, variant="secondary")
+                    with gr.Row():
+                        c_good = gr.Button("✅ 優質")
+                        c_risk = gr.Button("⚠️ 風險")
+                        c_pend = gr.Button("❓ 未定")
+                        c_rem = gr.Button("🗑️ 移除", variant="stop")
+
+            demo.load(comp_init, None, [comp_saved, c_df])
+            c_btn.click(comp_search, [c_in, comp_saved], [c_df, comp_res, c_load]).then(lambda: gr.update(value="搜尋結果"), outputs=[c_view])
+            c_load.click(comp_load_more, [c_in, comp_res, comp_saved], [c_df, comp_res])
+            c_view.change(comp_toggle, [c_view, comp_res, comp_saved], [c_df, c_load])
+            c_df.select(comp_select, [comp_res, comp_saved, c_view], [c_col, c_md, c_chat, comp_sel, comp_saved, c_tag_disp, c_tag_drop, c_tag_row])
+            c_send.click(comp_chat, [c_chat, c_msg, comp_sel], [c_chat, c_msg]); c_msg.submit(comp_chat, [c_chat, c_msg, comp_sel], [c_chat, c_msg])
+            c_tag_add.click(comp_add_tag, [c_tag_in, comp_sel, comp_saved, c_view, comp_res], [c_tag_in, c_tag_disp, c_tag_drop, comp_saved, c_df])
+            c_tag_del.click(comp_remove_tag, [c_tag_drop, comp_sel, comp_saved, c_view, comp_res], [c_tag_disp, c_tag_drop, comp_saved, c_df])
+            for btn, s in [(c_good,'good'),(c_risk,'risk'),(c_pend,'pending')]: btn.click(comp_update_status, [gr.State(s), comp_sel, comp_saved, c_view, comp_res], [c_df, comp_saved])
+            c_rem.click(comp_remove, [comp_sel, comp_saved, c_view, comp_res], [gr.State(None), c_df, comp_saved, c_col])
 
 if __name__ == "__main__":
-    main_app()
+    demo.launch()