postprocessing in correct order

matrix_pygmalion_bot/ai/koboldcpp.py

@@ -95,11 +95,10 @@ async def generate_sync(
 
     if complete_reply:
         complete_reply = complete_reply.removesuffix('<|endoftext|>')
-        complete_reply = complete_reply.replace(f"\n{bot.name}: ", " ")
-        complete_reply = complete_reply.replace(f"\n<BOT>: ", " ")
         complete_reply = complete_reply.replace(f"<BOT>", f"{bot.name}")
         complete_reply = complete_reply.replace(f"<USER>", f"You")
         complete_reply = complete_reply.replace(f"### Assistant", f"{bot.name}")
+        complete_reply = complete_reply.replace(f"\n{bot.name}: ", " ")
         return complete_reply.strip()
     else:
         raise ValueError(f"<ERROR> Timeout")

matrix_pygmalion_bot/ai/runpod_pygmalion.py

@@ -87,10 +87,9 @@ async def generate_sync(
                     reply = answer[:idx].strip()
                 else:
                     reply = answer.removesuffix('<|endoftext|>').strip()
-                reply = reply.replace(f"\n{bot.name}: ", " ")
-                reply = reply.replace(f"\n<BOT>: ", " ")
                 reply = reply.replace(f"<BOT>", f"{bot.name}")
                 reply = reply.replace(f"<USER>", f"You")
+                reply = reply.replace(f"\n{bot.name}: ", " ")
                 return reply
             else:
                 err_msg = r_json["error"] if "error" in r_json else ""