Max completion tokens only valid for threads

mylib/_chat.py

@@ -17,9 +17,11 @@ def parse_wait_time(err):
     raise TypeError(err.code)
 
 class ChatController:
-    _gpt_defaults = {
+    _assistant_kwargs = {
         'model': 'gpt-4o',
         'temperature': 1e-4,
+    }
+    _threads_kwargs = {
         'max_completion_tokens': 2 ** 12,
     }
 
@@ -28,7 +30,7 @@ class ChatController:
         self.database = database
         self.retries = retries
 
-        for i in self._gpt_defaults.items():
+        for i in self._assistant_kwargs.items():
             kwargs.setdefault(*i)
 
         self.assistant = self.client.beta.assistants.create(
@@ -73,6 +75,7 @@ class ChatController:
             run = self.client.beta.threads.runs.create_and_poll(
                 thread_id=self.thread.id,
                 assistant_id=self.assistant.id,
+                **self._threads_kwargs,
             )
             if run.status == 'completed':
                 return self.client.beta.threads.messages.list(