Update api/utils.py

api/utils.py

@@ -99,8 +99,8 @@ async def process_streaming_response(request: ChatRequest):
         "imageGenerationMode": False,
     }
 
-    response_content = ""
-    advertisement_added = False
+    response_content = ""  # Store accumulated response content
+    advertisement_added = False  # Track if advertisement is already added
 
     async with httpx.AsyncClient() as client:
         try:
@@ -126,10 +126,10 @@ async def process_streaming_response(request: ChatRequest):
                         # Add chunk to the response content
                         response_content += cleaned_content
 
-                        # Only yield cleaned content in chunks
+                        # Yield cleaned chunk as part of the stream (do not repeat this later)
                         yield f"data: {json.dumps(create_chat_completion_data(cleaned_content, request.model, timestamp))}\n\n"
 
-                # Append advertisement at the end, ensuring it happens only once
+                # Append advertisement at the very end of the response, once
                 if ADVERTISEMENT_TEXT and not advertisement_added:
                     response_content += "\n\n" + ADVERTISEMENT_TEXT
                     advertisement_added = True
@@ -191,7 +191,8 @@ async def process_non_streaming_response(request: ChatRequest):
         "imageGenerationMode": False,
     }
 
-    full_response = ""
+    full_response = ""  # Store full non-streaming response content
+
     async with httpx.AsyncClient() as client:
         try:
             async with client.stream("POST", f"{BASE_URL}/api/chat", headers=headers_api_chat, json=json_data) as response: