修复 chat()

app.py

@@ -31,32 +31,46 @@ print("Model loaded successfully!")
 
 def chat(message, history):
     """Handle chat interactions"""
-    # Build conversation (keep it short)
-    conversation = ""
-    
-    # Only use last 3 turns of history to keep context short
+
+    conversation = "System: You are a helpful assistant.\n"
+
+    # Only use last 3 turns
     recent_history = history[-3:] if len(history) > 3 else history
-    
-    for human, assistant in recent_history:
-        conversation += f"User: {human}\n"
-        conversation += f"Assistant: {assistant}\n"
-    
-    conversation += f"User: {message}\n"
-    conversation += "Assistant:"
-    
-    # Generate with optimized settings
+
+    for msg in recent_history:
+        # ChatInterface passes dict messages
+        if isinstance(msg, dict):
+            role = msg.get("role")
+            content = msg.get("content", "")
+
+            if role == "user":
+                conversation += f"User: {content}\n"
+            elif role == "assistant":
+                conversation += f"Assistant: {content}\n"
+
+        # fallback: old tuple/list history
+        else:
+            human, assistant = msg
+            conversation += f"User: {human}\n"
+            conversation += f"Assistant: {assistant or ''}\n"
+
+    # Add current user message
+    conversation += f"User: {message}\nAssistant:"
+
+    # LLM inference
     response = llm(
         conversation,
-        max_tokens=128,          # Reduced from 256 (faster)
+        max_tokens=128,
         temperature=0.7,
         top_p=0.9,
-        top_k=40,               # Added: limit sampling
+        top_k=40,
         repeat_penalty=1.1,
-        stop=["User:", "\n\n"],
-        echo=False,
+        stop=["User:", "Assistant:"],
+        echo=False
     )
-    
-    return response['choices'][0]['text'].strip()
+
+    return response["choices"][0]["text"].strip()
+
 
 # Create interface WITHOUT example caching
 demo = gr.ChatInterface(