Update chain_app.py

chain_app.py

@@ -2471,19 +2471,12 @@ async def main(message: cl.Message):
             stop=None,
         )
 
-        complete_content = ""
-
-        # Iterate over each chunk
         for chunk in completion:
-            # Retrieve the content from the current chunk
-            content = chunk.choices[0].delta.content
-            
+            # Retrieve the content from the current chunk            
             # Check if the content is not None before concatenating it
-            if content is not None:
-                complete_content += content
+            if chunk is not None:
+                await msg.stream_token(chunk.choices[0].delta.content)
 
-        # Send the concatenated content as a message
-        await cl.Message(content=complete_content).send()
 
     elif chat_profile == 'gemma-7B':
         completion = groq_client.chat.completions.create(
@@ -2501,20 +2494,12 @@ async def main(message: cl.Message):
             stop=None,
         )
 
-        complete_content = ""
-
-        # Iterate over each chunk
         for chunk in completion:
-            # Retrieve the content from the current chunk
-            content = chunk.choices[0].delta.content
-            
+            # Retrieve the content from the current chunk            
             # Check if the content is not None before concatenating it
-            if content is not None:
-                complete_content += content
+            if chunk is not None:
+                await msg.stream_token(chunk.choices[0].delta.content)
 
-        # Send the concatenated content as a message
-        await cl.Message(content=complete_content).send()
-    
     elif chat_profile == "zephyr-7B":
         result = hf_text_client.predict(
             message=message.content,
@@ -2524,10 +2509,8 @@ async def main(message: cl.Message):
             param_5=0.95,
             api_name="/chat"
         )
-        model_response = result.strip("</s>") 
-        await cl.Message(
-            content=model_response
-        ).send()
+        for token in result:
+            await msg.stream_token(token)
 
     elif chat_profile == 'mixtral-8x7B':
         completion = groq_client.chat.completions.create(