Update app.py

app.py

@@ -141,54 +141,59 @@ def respond(
     if not message.strip():
         return chat_history
 
-    system_message = f"""You are an interactive storyteller creating an immersive {genre or 'fantasy'} story.
-When responding:
-1. First, write 2-3 paragraphs (150-300 words) continuing the story based on the user's choice
-2. Use vivid sensory details and second-person perspective ("you", "your")
-3. Include dialogue and character thoughts
-4. After the story segment, provide exactly three numbered choices for what to do next
-5. Format the choices as:
+    # Use the enhanced system prompt instead of the simple one
+    system_message = get_enhanced_system_prompt(genre)
 
-1. You [action verb]...
-2. You [action verb]...
-3. You [action verb]...
-"""
-
-    # Format history properly for the API
-    formatted_messages = [{"role": "system", "content": system_message}]
+    # Format messages for the API
+    formatted_messages = []
+    
+    # Add system message first
+    formatted_messages.append({
+        "role": "system",
+        "content": system_message
+    })
     
-    if chat_history:
+    # Add chat history
+    if chat_history and use_full_memory:
         for user_msg, bot_msg in chat_history[-MEMORY_WINDOW:]:
             formatted_messages.extend([
                 {"role": "user", "content": str(user_msg)},
                 {"role": "assistant", "content": str(bot_msg)}
             ])
     
-    formatted_messages.append({"role": "user", "content": str(message)})
-    
+    # Add current message
+    formatted_messages.append({
+        "role": "user", 
+        "content": str(message)
+    })
+
     try:
         bot_message = ""
+        current_history = list(chat_history)
+        
         for response_chunk in client.chat_completion(
             formatted_messages,
             max_tokens=MAX_TOKENS,
             stream=True,
             temperature=TEMPERATURE,
-            top_p=TOP_P,
+            top_p=TOP_P
         ):
             if hasattr(response_chunk.choices[0].delta, 'content'):
                 delta = response_chunk.choices[0].delta.content
                 if delta:
                     bot_message += delta
                     if len(bot_message.strip()) >= MIN_RESPONSE_LENGTH:
-                        chat_history = list(chat_history)
-                        chat_history.append((message, bot_message))
-                        yield chat_history
+                        new_history = current_history + [(message, bot_message)]
+                        yield new_history
+        
+        if bot_message:  # Ensure final message is yielded
+            final_history = current_history + [(message, bot_message)]
+            yield final_history
                         
     except Exception as e:
         error_message = f"Story magic temporarily interrupted. Please try again. (Error: {str(e)})"
-        chat_history = list(chat_history)
-        chat_history.append((message, error_message))
-        yield chat_history
+        error_history = current_history + [(message, error_message)]
+        yield error_history
 
 def save_story(chat_history):
     """Convert chat history to markdown for download"""
@@ -279,7 +284,7 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
             # 4) Connect each starter button:
             for starter_button in starter_buttons:
                 starter_button.click(
-                    fn=lambda x: ((x, ""), []), # Show user message immediately
+                    fn=lambda x: [(x, "")],  # Changed to return properly formatted history
                     inputs=[starter_button],
                     outputs=[chatbot],
                     queue=False