Update app.py

app.py

@@ -5,92 +5,74 @@ from langchain.text_splitter import RecursiveCharacterTextSplitter
 from langchain_community.vectorstores import FAISS
 from langchain_huggingface import HuggingFaceEmbeddings
 from docx import Document as DocxDocument
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 from huggingface_hub import login, snapshot_download
 import gradio as gr
 
 # -------------------------------
-# 1. 模型設定（中文 T5 + fallback）
+# 1. 模型設定（專門中文，T5）
 # -------------------------------
-PRIMARY_MODEL = "Langboat/mengzi-t5-base"  # ✅ 帶 spiece.model
-FALLBACK_MODEL = "uer/t5-small-chinese-cluecorpussmall"
+MODEL_NAME = "Langboat/mengzi-t5-base"  # ✅ CPU 也能跑的中文 T5
+LOCAL_MODEL_DIR = f"./models/{MODEL_NAME.split('/')[-1]}"
 
 HF_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
 if HF_TOKEN:
     login(token=HF_TOKEN)
     print("✅ 已使用 HUGGINGFACEHUB_API_TOKEN 登入 Hugging Face")
 
-def try_download_model(repo_id):
-    local_dir = f"./models/{repo_id.split('/')[-1]}"
-    if not os.path.exists(local_dir):
-        print(f"⬇️ 嘗試下載模型 {repo_id} ...")
-        try:
-            snapshot_download(repo_id=repo_id, token=HF_TOKEN, local_dir=local_dir)
-        except Exception as e:
-            print(f"⚠️ 模型 {repo_id} 無法下載: {e}")
-            return None
-    return local_dir
-
-# 嘗試下載 Primary，失敗就換 Small
-LOCAL_MODEL_DIR = try_download_model(PRIMARY_MODEL)
-if LOCAL_MODEL_DIR is None:
-    print("⚠️ 切換到 fallback 模型：小型 T5-Chinese")
-    LOCAL_MODEL_DIR = try_download_model(FALLBACK_MODEL)
-    MODEL_NAME = FALLBACK_MODEL
-else:
-    MODEL_NAME = PRIMARY_MODEL
+if not os.path.exists(LOCAL_MODEL_DIR):
+    print(f"⬇️ 嘗試下載模型 {MODEL_NAME} ...")
+    snapshot_download(repo_id=MODEL_NAME, token=HF_TOKEN, local_dir=LOCAL_MODEL_DIR)
 
 print(f"👉 最終使用模型：{MODEL_NAME}")
 
 # -------------------------------
-# 2. pipeline 載入 (Seq2SeqLM for T5)
+# 2. 載入 tokenizer + model
 # -------------------------------
 tokenizer = AutoTokenizer.from_pretrained(LOCAL_MODEL_DIR)
-model = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_DIR)
-
-generator = pipeline(
-    "text2text-generation",
-    model=model,
-    tokenizer=tokenizer,
-    device=-1  # CPU
-)
-
-def call_local_inference(prompt, max_new_tokens=256):
-    try:
-        outputs = generator(
-            prompt,
-            max_new_tokens=max_new_tokens,
-            do_sample=True,
-            temperature=0.7
-        )
-        return outputs[0]["generated_text"]
-    except Exception as e:
-        return f"（生成失敗：{e}）"
+model = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_DIR, device_map="cpu")
 
 # -------------------------------
-# 3. 建立/載入向量資料庫
+# 3. 向量資料庫載入
 # -------------------------------
 EMBEDDINGS_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
 embeddings_model = HuggingFaceEmbeddings(model_name=EMBEDDINGS_MODEL_NAME)
 
-DB_PATH = "./faiss_db"
-if os.path.exists(os.path.join(DB_PATH, "index.faiss")):
+if os.path.exists("./faiss_db/index.faiss"):
     print("✅ 載入現有向量資料庫...")
-    db = FAISS.load_local(DB_PATH, embeddings_model, allow_dangerous_deserialization=True)
+    db = FAISS.load_local("./faiss_db", embeddings_model, allow_dangerous_deserialization=True)
 else:
-    print("⚠️ 沒有找到資料庫，請先建立 faiss_db")
+    print("⚠️ 找不到向量資料庫，請先建立")
     db = None
 
-retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 3}) if db else None
+retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 5}) if db else None
 
 # -------------------------------
-# 4. 文章生成（加入 RAG）
+# 4. 改良推理函數（避免重複亂碼）
+# -------------------------------
+def call_local_inference(prompt, max_new_tokens=256):
+    try:
+        inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(model.device)
+        outputs = model.generate(
+            **inputs,
+            max_new_tokens=max_new_tokens,
+            do_sample=False,            # ❌ 關掉隨機
+            num_beams=4,                # ✅ 用 beam search
+            repetition_penalty=1.2,     # ✅ 懲罰重複
+            length_penalty=1.0,
+            early_stopping=True
+        )
+        return tokenizer.decode(outputs[0], skip_special_tokens=True)
+    except Exception as e:
+        return f"（生成失敗：{e}）"
+
+# -------------------------------
+# 5. 文章生成（加入 RAG）
 # -------------------------------
 def generate_article_progress(query, segments=5):
     docx_file = "/tmp/generated_article.docx"
     doc = DocxDocument()
     doc.add_heading(query, level=1)
-
     all_text = []
 
     context = ""
@@ -101,30 +83,34 @@ def generate_article_progress(query, segments=5):
 
     for i in range(segments):
         prompt = (
-            f"以下是佛教經論的相關段落：\n{context}\n\n"
-            f"請依據上面內容，寫一段約150-200字的中文文章，"
-            f"主題：{query}。\n第{i+1}段："
+            f"請基於以下資料，撰寫一段中文文章：\n"
+            f"主題：{query}\n"
+            f"要求：字數約150~200字，內容要有完整句子，不要重複詞語。\n\n"
+            f"參考資料：\n{context}\n\n"
+            f"第{i+1}段："
         )
+
         paragraph = call_local_inference(prompt)
         all_text.append(paragraph)
         doc.add_paragraph(paragraph)
+
         yield "\n\n".join(all_text), None, f"本次使用模型：{MODEL_NAME}"
 
     doc.save(docx_file)
     yield "\n\n".join(all_text), docx_file, f"本次使用模型：{MODEL_NAME}"
 
 # -------------------------------
-# 5. Gradio 介面
+# 6. Gradio 介面
 # -------------------------------
 with gr.Blocks() as demo:
     gr.Markdown("# 📺 電視弘法視頻生成文章 RAG 系統")
-    gr.Markdown("使用 FAISS + 中文 T5 模型，基於資料庫內容生成文章。")
+    gr.Markdown("基於向量資料庫 + 中文 T5 模型，自動生成主題文章")
 
     query_input = gr.Textbox(lines=2, placeholder="請輸入文章主題", label="文章主題")
-    segments_input = gr.Slider(minimum=1, maximum=10, step=1, value=5, label="段落數")
+    segments_input = gr.Slider(minimum=1, maximum=10, step=1, value=3, label="段落數")
     output_text = gr.Textbox(label="生成文章")
     output_file = gr.File(label="下載 DOCX")
-    model_info = gr.Label(label="模型資訊")
+    model_info = gr.Textbox(label="模型資訊")
 
     btn = gr.Button("生成文章")
     btn.click(