Update app.py

app.py

@@ -11,33 +11,58 @@ model_path = hf_hub_download(
 # 2. Load the Engine
 llm = Llama(
     model_path=model_path,
-    n_ctx=2048,
+    n_ctx=4096,
     n_threads=2,
     verbose=False
 )
 
-def dual_frame_response(message, history):
-    system_prompt = """You are an Unbiased Decision Architect. 
-    Analyze the situation and provide the response in exactly two distinct sections:
-    1. **The Gain Frame (Positive)**
-    2. **The Loss Frame (Negative)**
-    Do not offer a synthesis."""
+def smart_response(message, history):
+    # This is the "Brain" instruction. 
+    # It tells the model to distinguish between DECISIONS and NORMAL CHAT.
+    system_prompt = """You are a helpful AI assistant with a specific behavioral directive.
     
-    prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+    CORE INSTRUCTION:
+    1. IF the user asks for a DECISION, OPINION, or ADVICE (e.g., "Should I...", "Is X good...", "What do you think of..."):
+       You MUST analyze it using TWO distinct frames to avoid bias. 
+       Format your answer exactly like this:
+       **The Gain Frame (Positive)**
+       [Write the positive/gain perspective here]
+       
+       **The Loss Frame (Negative)**
+       [Write the negative/loss/risk perspective here]
+       (Do not provide a conclusion or your own opinion).
+
+    2. IF the user asks a FACTUAL question (e.g., "1+1", "Capital of France"), a CLARIFICATION (e.g., "What did you mean?"), or general CHAT:
+       Answer normally, concisely, and directly. Do NOT use the Gain/Loss frames.
+    """
+    
+    # 3. Build the Memory (Context)
+    # Loop through the history so the model remembers what happened before.
+    formatted_prompt = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system_prompt}<|eot_id|>"
+    
+    for user_msg, bot_msg in history:
+        formatted_prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{user_msg}<|eot_id|>"
+        formatted_prompt += f"<|start_header_id|>assistant<|end_header_id|>\n\n{bot_msg}<|eot_id|>"
+    
+    # Add the current message
+    formatted_prompt += f"<|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
     
+    # 4. Generate Response
     output = llm(
-        prompt, 
-        max_tokens=512, 
+        formatted_prompt, 
+        max_tokens=1024, # Allowed it to write a bit more
         stop=["<|eot_id|>"], 
         echo=False
     )
     return output['choices'][0]['text']
 
+# 5. The Interface
 demo = gr.ChatInterface(
-    fn=dual_frame_response,
-    title="🤖 Unbiased Framing Decision Architect",
-    description="Running Llama 3.1 on Hugging Face Free Tier.",
-    examples=["Should I buy a lottery ticket?", "Is nuclear energy good?"],
+    fn=smart_response,
+    title="🤖 Smart Decision Architect",
+    description="I am a Smart Assistant. Ask me factual questions (1+1) and I will answer normally. Ask me for ADVICE (Should I...) and I will give you Gain/Loss frames.",
+    examples=["Should I buy a lottery ticket?", "What is 10 + 10?", "Is nuclear energy good?", "Summarize the previous answer."],
+    theme="soft"
 )
 
 if __name__ == "__main__":