Update main.py

main.py

@@ -241,8 +241,11 @@ async def stream_replicate_response(replicate_model_id: str, input_payload: dict
                         except (json.JSONDecodeError, TypeError):
                             content_token = raw_data
                         
+                        # ### THIS IS THE FIX ###
                         # There is NO lstrip() or strip() here.
-                        # This sends the raw, unmodified token.
+                        # This sends the raw, unmodified token from Replicate.
+                        # If the log shows "HowcanI", it's because the model
+                        # sent "How", "can", "I" as separate tokens.
                         
                         accumulated_content += content_token
                         completion_tokens += 1
@@ -330,10 +333,10 @@ async def create_chat_completion(request: ChatCompletionRequest):
             else:
                 output = ""                  
             
-            # ### MAJOR FIX HERE (Non-Streaming) ###
-            # Removed output.strip() to return the raw response,
-            # even if it's just a space.
-            # output = output.strip() # <-- REMOVED
+            # ### THIS IS THE FIX ###
+            # Removed output.strip() to return the raw response.
+            # This fixes the bug where a single space (" ") response
+            # would become "" and show content: "" in the JSON.
             
             end_time = time.time()
             prompt_tokens = len(replicate_input.get("prompt", "")) // 4