Will chat work better with system prompt?

app.py

@@ -35,11 +35,11 @@ model_info ={
 }
 
 def format_promt(message, custom_instructions=None):
-    prompt = ""
+    messages = []
     if custom_instructions:
-        prompt += f"system\n{custom_instructions}\n"
-    prompt += f"user\n{message}\n"
-    return prompt
+        messages.append({"role": "system", "content": custom_instructions})
+    messages.append({"role": "user", "content": message})
+    return {"inputs": {"messages": messages}}
 
 def reset_conversation():
     '''
@@ -79,7 +79,7 @@ temp_values = st.sidebar.slider(
 
 custom_instructions = st.sidebar.text_area(
     "Custom Instructions",
-    value="You are helpful assistant, act like a Human in conversation. Keep asnwers short and in English only!",
+    value="You are helpful assistant, act like a Human in conversation. Keep answers very short and in English only!",
     help="Customize how the AI should behave"
 )
 
@@ -130,11 +130,13 @@ if prompt := st.chat_input(f"Hi I'm {selected_model}, How can I help you today?"
         client = InferenceClient(
             model=model_links[selected_model],)
 
-        output = client.text_generation(
-            formated_text,
-            temperature=temp_values,#0.5
-            max_new_tokens=1000,
-            stream=True
+        output = client.post(
+            json=formated_text,
+            params={
+                "temperature": temp_values,
+                "max_new_tokens": 1000,
+                "stream": True
+            }
         )
 
         # Create a placeholder for the streaming response
@@ -143,7 +145,14 @@ if prompt := st.chat_input(f"Hi I'm {selected_model}, How can I help you today?"
 
         # Stream the response and accumulate it
         for chunk in output:
-            full_response += chunk
+            if isinstance(chunk, dict) and "generated_text" in chunk:
+                text_chunk = chunk["generated_text"]
+            elif isinstance(chunk, str):
+                text_chunk = chunk
+            else:
+                continue
+                
+            full_response += text_chunk
             message_placeholder.markdown(full_response + "▌")
         
         # Display final response and store it