Update app.py

app.py

@@ -21,11 +21,13 @@ llm = Llama(
 def smart_response(message, history):
     
     # --- SMART ROUTER ---
-    triggers = ["should", "opinion", "think", "good", "bad", "pros", "cons", "benefit", "risk", "impact", "why"]
+    # We check if the user is asking for an opinion/analysis vs a simple fact
+    triggers = ["should", "opinion", "think", "good", "bad", "pros", "cons", "benefit", "risk", "impact", "why", "compare", "difference"]
     is_opinion = any(t in message.lower() for t in triggers)
 
     if is_opinion:
         print(f"🧠 OPINION MODE: {message}")
+        # Use the format the model was fine-tuned on to trigger the Dual behavior
         prompt = f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
 
 ### Instruction:
@@ -35,28 +37,37 @@ def smart_response(message, history):
 """
     else:
         print(f"ℹ️ CHAT MODE: {message}")
-        # Even in chat mode, we use the formatting the model knows best to prevent hallucinations
+        # STRONG OVERRIDE: We explicitly tell the model strictly NOT to use framing.
         prompt = f"""Below is an instruction that describes a task. Write a response that appropriately completes the request.
 
 ### Instruction:
-Answer the following question concisely: {message}
+You are a factual assistant. Answer the following question directly and concisely. Do NOT use positive or negative framing.
+Question: {message}
 
 ### Response:
 """
 
-    # --- GENERATION WITH STREAMING ---
+    # --- GENERATION ---
     stream = llm(
         prompt,
         max_tokens=512,
         temperature=0.7,
-        stop=["</s>", "###", "Instruction:", "[INST]"],
-        stream=True,
+        stop=["</s>", "###", "Instruction:"],
+        stream=True, 
         echo=False
     )
     
     partial_message = ""
     for output in stream:
         token = output['choices'][0]['text']
+        
+        # --- SAFETY NET ---
+        # If the model tries to output the headers in Chat Mode, we hide them.
+        # This acts as a filter for the "overfitting" behavior.
+        if not is_opinion:
+            if "**POSITIVE FRAMING:**" in token or "**NEGATIVE FRAMING:**" in token:
+                continue # Skip printing these specific tokens
+                
         partial_message += token
         yield partial_message