sys prompt

app.py

@@ -78,33 +78,43 @@ def add_docs(user_id: str, docs: list[str]) -> int:
     return len(docs)
 # ----- Qwen-chat prompt helper ---------------------------------------------
 def build_llm_prompt(system: str, context: list[str], user_question: str) -> str:
-    """Return a Qwen-style prompt with multiple context items."""
+    """
+    建立適用於 LLaMA/Qwen 等模型的 prompt，支援多段 context，
+    並強化 system prompt 限制模型僅輸出回應內容。
+    """
     load_chat()  # 確保 tokenizer 載入
 
+    # 強化指令：防止解釋與步驟
+    system_prompt = (
+        f"{system.strip()}\n"
+        "Do not include any explanations, steps, or analysis. "
+        "Only output the final reply content."
+    )
+
     conversation = [
-        {"role": "system", "content": system}
+        {"role": "system", "content": system_prompt}
     ]
 
-    # 將每段 context 當作 user 提供的提示
+    # 多段 context 當作 user 發言
     for ctx in context:
-        if ctx.strip():  # 跳過空內容
-            conversation.append({"role": "user", "content": ctx})
+        if ctx.strip():  # 忽略空內容
+            conversation.append({"role": "user", "content": ctx.strip()})
 
-    # 加入最終問題
-    conversation.append({"role": "user", "content": user_question})
+    # 最後加入使用者問題
+    conversation.append({"role": "user", "content": user_question.strip()})
 
+    # 套用 LLaMA-style prompt 格式
     prompt = ""
     for turn in conversation:
         role = turn["role"]
-        content = turn["content"].strip()
+        content = turn["content"]
         if role == "system":
             prompt += f"<<SYS>>\n{content}\n<</SYS>>\n\n"
         elif role == "user":
-            prompt += f"[INST] {content.strip()} [/INST]\n"
+            prompt += f"[INST] {content} [/INST]\n"
         elif role == "assistant":
-            prompt += f"{content.strip()}\n"
+            prompt += f"{content}\n"
     return prompt
-    
 
 # ---------- 4. Gradio playground (same UI as before) --------------------------
 def store_doc(doc_text: str, user_id="demo"):