app.py

import gradio as gr
import pandas as pd
from sentence_transformers import SentenceTransformer, util
import torch
import os
import sys
import gc 
import time

# --- 系統設定 ---
SYSTEM_TITLE = "花蓮慈濟醫院公文輔助判決系統"
FILE_PATH = 'data.csv'
INDEX_FILE = 'corpus_embeddings.pt' 

# ▼▼▼ 設定登入帳號密碼 (已更新) ▼▼▼
# 格式：("帳號", "密碼")
LOGIN_DATA = ("admin", "htch15583") 

# --- 1. 讀取資料 ---
print("🚀 正在啟動系統...")

if not os.path.exists(FILE_PATH):
    print(f"❌ 錯誤：找不到 {FILE_PATH}")
    sys.exit(1)

try:
    # 讀取檔案 (CP950 優先)
    df = pd.read_csv(FILE_PATH, encoding='cp950')
except UnicodeDecodeError:
    try:
        df = pd.read_csv(FILE_PATH, encoding='big5')
    except Exception:
        df = pd.DataFrame()
except Exception:
    df = pd.DataFrame()

# --- 2. 資料清洗 ---
if not df.empty:
    # 移除 BOM 與空白
    df.columns = [str(c).strip().replace('\ufeff', '') for c in df.columns]
    
    # 自動對應欄位
    for col in df.columns:
        if '主旨' in col or '內容' in col: df.rename(columns={col: '主旨'}, inplace=True)
        if '窗口' in col or '單位' in col: df.rename(columns={col: '收文窗口'}, inplace=True)

    # 轉字串 & 移除空值
    df['主旨'] = df['主旨'].astype(str)
    df['收文窗口'] = df['收文窗口'].astype(str)
    df = df.dropna(subset=['主旨', '收文窗口'])
    
    corpus = df['主旨'].tolist()
    total_records = len(corpus)
    print(f"📊 載入全量資料: {total_records} 筆")
else:
    print("❌ 資料表是空的！")
    corpus = []
    total_records = 0

# --- 3. 載入模型與建立索引 ---
model = None
try:
    print("🧠 正在載入模型 (BAAI/bge-small-zh-v1.5)...")
    model = SentenceTransformer('BAAI/bge-small-zh-v1.5')
except Exception as e:
    print(f"❌ 模型載入失敗: {e}")

corpus_embeddings = None

if total_records > 0 and model is not None:
    # 檢查是否有快取檔案
    if os.path.exists(INDEX_FILE):
        print(f"⚡ 偵測到快取檔案，正在秒速載入...")
        try:
            corpus_embeddings = torch.load(INDEX_FILE)
            print("✅ 索引載入完成！")
        except Exception as e:
            print(f"❌ 快取檔案損壞，將重新計算。錯誤: {e}")
            corpus_embeddings = None
    
    # 如果沒有快取，則進行計算
    if corpus_embeddings is None:
        print(f"🔥 開始計算索引 (需時約 2-4 分鐘，請耐心等候)...")
        chunk_size = 500 
        embeddings_chunks = []
        
        try:
            for i in range(0, total_records, chunk_size):
                batch = corpus[i : i + chunk_size]
                batch_emb = model.encode(batch, convert_to_tensor=True, show_progress_bar=False)
                embeddings_chunks.append(batch_emb)
                print(f"   -> 已處理 {min(i + chunk_size, total_records)} / {total_records} 筆...")
                gc.collect() 
            
            corpus_embeddings = torch.cat(embeddings_chunks)
            # 儲存到硬碟，下次啟動就會很快
            torch.save(corpus_embeddings, INDEX_FILE)
            print("✅ 索引計算並儲存完成！")
            
        except Exception as e:
            print(f"❌ 索引計算失敗: {e}")
            corpus_embeddings = None

# --- 4. 定義搜尋 ---
def search_department(query):
    if corpus_embeddings is None:
        return "⚠️ 系統初始化失敗，請檢查 Logs。"
    
    if not query.strip():
        return "請輸入公文主旨..."

    query_embedding = model.encode(query, convert_to_tensor=True)
    cos_scores = util.cos_sim(query_embedding, corpus_embeddings)[0]
    
    top_k = min(3, len(corpus))
    top_results = torch.topk(cos_scores, k=top_k)
    
    output_text = f"🔍 分析結果：\n" + "="*30 + "\n"
    
    for score, idx in zip(top_results.values, top_results.indices):
        idx = idx.item()
        if idx < len(df):
            row = df.iloc[idx]
            score_val = score.item()
            
            if score_val > 0.7: confidence = "⭐⭐⭐ 極高"
            elif score_val > 0.55: confidence = "⭐⭐ 高"
            else: confidence = "⭐ 參考"
            
            output_text += f"【推薦單位】：{row['收文窗口']}\n"
            output_text += f"  - 歷史案例：{row['主旨']}\n"
            output_text += f"  - 相似度：{score_val:.4f} ({confidence})\n"
            output_text += "-"*20 + "\n"
            
    return output_text

# --- 5. 介面 (已啟用密碼鎖) ---
iface = gr.Interface(
    fn=search_department,
    inputs=gr.Textbox(lines=3, placeholder="請輸入公文主旨..."),
    outputs=gr.Textbox(lines=12, label="AI 判決建議"),
    title=SYSTEM_TITLE,
    description=f"系統狀態：{'🟢 系統正常' if corpus_embeddings is not None else '🔴 異常'}\n資料庫收錄：{total_records} 筆歷史資料",
    examples=[["檢送本署彙整人工生殖機構之捐贈生殖細胞使用情形"], ["函轉衛生局關於流感疫苗接種計畫"]]
)

if __name__ == "__main__":
    # 啟動時加入驗證
    iface.launch(auth=LOGIN_DATA)