Update api/utils.py

api/utils.py

@@ -101,7 +101,6 @@ async def process_streaming_response(request: ChatRequest):
         "imageGenerationMode": False,
     }
 
-    response_content = ""  # Store accumulated response content
     advertisement_added = False  # Track if advertisement is already added
 
     async with httpx.AsyncClient() as client:
@@ -125,19 +124,16 @@ async def process_streaming_response(request: ChatRequest):
 
                         cleaned_content = strip_model_prefix(content, model_prefix)
 
-                        # Add chunk to the response content
-                        response_content += cleaned_content
-
-                        # Yield cleaned chunk as part of the stream (do not repeat this later)
+                        # Yield cleaned chunk as part of the stream
                         yield f"data: {json.dumps(create_chat_completion_data(cleaned_content, request.model, timestamp))}\n\n"
 
-                # Append advertisement at the very end of the response, once
+                # Append advertisement as a separate chunk, if applicable
                 if ADVERTISEMENT_TEXT and not advertisement_added:
-                    response_content += "\n\n" + ADVERTISEMENT_TEXT
+                    yield f"data: {json.dumps(create_chat_completion_data(ADVERTISEMENT_TEXT, request.model, timestamp))}\n\n"
                     advertisement_added = True
 
-                # Final chunk with complete content
-                yield f"data: {json.dumps(create_chat_completion_data(response_content, request.model, timestamp, 'stop'))}\n\n"
+                # Yield final chunk indicating the end of the stream
+                yield f"data: {json.dumps(create_chat_completion_data('', request.model, timestamp, 'stop'))}\n\n"
                 yield "data: [DONE]\n\n"
 
         except httpx.HTTPStatusError as e: