Update app.py

app.py

@@ -10,10 +10,10 @@ from huggingface_hub import login, snapshot_download
 import gradio as gr
 
 # -------------------------------
-# 1. 模型設定（中文 T5 / Pegasus）
+# 1. 模型設定（中文 T5 + fallback）
 # -------------------------------
-PRIMARY_MODEL = "imxly/t5-pegasus-small"   # 適合中文摘要/生成
-FALLBACK_MODEL = "uer/gpt2-chinese-cluecorpussmall"  # 若 T5 無法下載就 fallback GPT2
+PRIMARY_MODEL = "Langboat/mengzi-t5-base"  # ✅ 帶 spiece.model
+FALLBACK_MODEL = "uer/t5-small-chinese-cluecorpussmall"
 
 HF_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
 if HF_TOKEN:
@@ -31,9 +31,10 @@ def try_download_model(repo_id):
             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 GPT2-Chinese")
+    print("⚠️ 切換到 fallback 模型：小型 T5-Chinese")
     LOCAL_MODEL_DIR = try_download_model(FALLBACK_MODEL)
     MODEL_NAME = FALLBACK_MODEL
 else:
@@ -42,28 +43,16 @@ else:
 print(f"👉 最終使用模型：{MODEL_NAME}")
 
 # -------------------------------
-# 2. pipeline 載入
+# 2. pipeline 載入 (Seq2SeqLM for T5)
 # -------------------------------
-tokenizer = AutoTokenizer.from_pretrained(
-    LOCAL_MODEL_DIR,
-    use_fast=False  # 防止 sentencepiece 問題
-)
-
-# 判斷 GPU (CL3) 或 CPU
-device = 0 if torch.cuda.is_available() else -1
-print(f"💻 使用裝置：{'GPU' if device == 0 else 'CPU'}")
-
-try:
-    model = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_DIR)
-except:
-    from transformers import AutoModelForCausalLM
-    model = AutoModelForCausalLM.from_pretrained(LOCAL_MODEL_DIR)
+tokenizer = AutoTokenizer.from_pretrained(LOCAL_MODEL_DIR)
+model = AutoModelForSeq2SeqLM.from_pretrained(LOCAL_MODEL_DIR)
 
 generator = pipeline(
-    "text2text-generation" if "t5" in MODEL_NAME or "pegasus" in MODEL_NAME else "text-generation",
+    "text2text-generation",
     model=model,
     tokenizer=tokenizer,
-    device=device
+    device=-1  # CPU
 )
 
 def call_local_inference(prompt, max_new_tokens=256):
@@ -79,68 +68,69 @@ def call_local_inference(prompt, max_new_tokens=256):
         return f"（生成失敗：{e}）"
 
 # -------------------------------
-# 3. FAISS 向量資料庫載入
+# 3. 建立/載入向量資料庫
 # -------------------------------
-DB_PATH = "./faiss_db"
 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")):
     print("✅ 載入現有向量資料庫...")
     db = FAISS.load_local(DB_PATH, embeddings_model, allow_dangerous_deserialization=True)
 else:
-    print("⚠️ 找不到向量資料庫，將建立空的 DB")
-    db = FAISS.from_documents([], embeddings_model)
+    print("⚠️ 沒有找到資料庫，請先建立 faiss_db")
+    db = None
 
-retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 5})
+retriever = db.as_retriever(search_type="similarity", search_kwargs={"k": 3}) if db else None
 
 # -------------------------------
-# 4. 文章生成（結合 RAG）
+# 4. 文章生成（加入 RAG）
 # -------------------------------
-def generate_article_progress(query, segments=3):
+def generate_article_progress(query, segments=5):
     docx_file = "/tmp/generated_article.docx"
     doc = DocxDocument()
     doc.add_heading(query, level=1)
 
     all_text = []
 
-    retrieved_docs = retriever.get_relevant_documents(query)
-    context_texts = [d.page_content for d in retrieved_docs]
-    context = "\n".join([f"{i+1}. {txt}" for i, txt in enumerate(context_texts[:3])])
+    context = ""
+    if retriever:
+        retrieved_docs = retriever.get_relevant_documents(query)
+        context_texts = [d.page_content for d in retrieved_docs]
+        context = "\n".join([f"{i+1}. {txt}" for i, txt in enumerate(context_texts[:3])])
 
     for i in range(segments):
         prompt = (
-            f"以下是佛教經論的相關內容：\n{context}\n\n"
+            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："
         )
         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}，裝置：{'GPU' if device == 0 else 'CPU'}"
+        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}，裝置：{'GPU' if device == 0 else 'CPU'}"
+    yield "\n\n".join(all_text), docx_file, f"本次使用模型：{MODEL_NAME}"
 
 # -------------------------------
 # 5. 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 Hugging Face 本地模型 + FAISS RAG，僅基於資料庫生成文章。")
+    gr.Markdown("使用 FAISS + 中文 T5 模型，基於資料庫內容生成文章。")
 
     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=3, label="段落數")
+    segments_input = gr.Slider(minimum=1, maximum=10, step=1, value=5, label="段落數")
     output_text = gr.Textbox(label="生成文章")
     output_file = gr.File(label="下載 DOCX")
-    status_info = gr.Label(label="狀態")
+    model_info = gr.Label(label="模型資訊")
 
     btn = gr.Button("生成文章")
     btn.click(
         generate_article_progress,
         inputs=[query_input, segments_input],
-        outputs=[output_text, output_file, status_info]
+        outputs=[output_text, output_file, model_info]
     )
 
 if __name__ == "__main__":