Fixed Streaming response being double wrapped

main/routes.py

@@ -74,6 +74,13 @@ async def create_chat_completion(request: ChatCompletionRequest):
                 async for chunk in api.generate_stream(
                         prompt=last_message,
                 ):
+                    # Parse the SSE format from LLM Server
+                    if chunk.startswith('data: '):
+                        chunk = chunk[6:].strip()  # Remove "data: " and trailing \n\n
+
+                    if chunk == '[DONE]':
+                        continue
+
                     response_chunk = {
                         "id": "chatcmpl-123",
                         "object": "chat.completion.chunk",
@@ -87,7 +94,6 @@ async def create_chat_completion(request: ChatCompletionRequest):
                             "finish_reason": None
                         }]
                     }
-                    # Need to format this exactly as SSE requires
                     yield f"data: {json.dumps(response_chunk)}\n\n"
                 yield "data: [DONE]\n\n"