All my runs are "Status: incomplete"

A few minutes ago I was testing code just fine with no problem and suddenly all the runs I submit immediately go into “incomplete” status. Is this an outage of sorts, or what would cause this? I have plenty of credits in my account and run.last_error is None.

It has been going on for the better part of 30-45 minutes.

2 Likes

Now receiving run failed errors and this:

Last error was LastError(code=‘server_error’, message=‘Sorry, something went wrong.’) at 1717444364.

Is anyone from OpenAI monitoring these forums that can confirm if there is an issue on their side?

1 Like

@satoshin - Welcome to the forum,

You can find issues and recent incidents below. Hope this helps
https://status.openai.com/

2 Likes

Hello @satoshin , I am facing the same issue it gets stuck in incomplete status for long time and then in some response directly gives
Last error was LastError(code=‘server_error’, message=‘Sorry, something went wrong.’)

I am stuck since last week do you have any updates on same ?

What I notice is, if a run status goes to incomplete, it never recovers. The only option seems to be to start a new thread :frowning:

What does “incomplete” even mean - is there any documentation on what went wrong and how to recover?

Found a piece of documentation which could probably throw more light on what happened - https://platform.openai.com/docs/api-reference/runs/object#runs/object-status

The run.incomplete_details property on the run object.

Yeah the reason was [quote=“satoshin, post:2, topic:794185”]
Sorry, something went wrong.
[/quote]

lol. It now magically works again. I guess it’s a mystery for now.

If you have max tokens populated with a value try removing it and see if you get different results.

1 Like

@satoshin Thanks let me check if that solves the issue for now

I tried with different max_prompt_tokens and max_completion_tokens from null to 50000. Yet, kept getting an incomplete run.

The issue resolved for the particular thread, when I set the truncation_strategy to last_messages
And the issue would re-surface on the same thread, if I let the truncation_strategy take the default i.e. auto.

The incomplete reason, though is always -
incomplete_details: {"reason":"max_prompt_tokens"}

Overall, the behaviour seems more stable with the truncation_strategy of last_messages, though I have seen an incomplete run once even with this strategy :face_with_open_eyes_and_hand_over_mouth:

1 Like

@jknt Agree I did the same strategy, when you deal with max tokens it always stuck
with incomplete status and if truncation_strategy is auto then also it keeps failing after certain questions while setting last_messages makes the thread stable

@satoshin Thank you for the solution Removing Max tokens helped but then again what is the use of the parameter if we cant control the flow

The reason is: using more tokens makes OpenAi more money. lol.

I noticed that they even start throttling ChatGPT for me after a couple of hours of using it with coding help since I’m not a professional developer.

1 Like

Hey @satoshin!

Happy to help with this, Can you provide me with a thread id where you’re seeing this behavior?

1 Like

@stevecoffey Here is a threadId which is seeing this error:
thread_U0xzyGVHXEyOGGzCnpZAIU3Z

and the run is created like this:

 const run = await openai.beta.threads.runs.create(threadId, {
        assistant_id: assistantId,
        max_prompt_tokens: 50000,
        max_completion_tokens: 50000,
        truncation_strategy: {
          type: 'last_messages',
          last_messages: 10,
        },
      });

I tried with truncation_strategy of type auto and even that fails after a few back and forth.

If I create a new thread, all starts working again for a little while.

@stevecoffey Okay on a little more digging, I see that while the API gives the error, internally the model seems to have generated 11 replies to a single prompt of hi. Please check the last set of exchanges on the thread.

The replies can be viewed on the thread https://platform.openai.com/threads/thread_U0xzyGVHXEyOGGzCnpZAIU3Z

Whereas the API continues to return an error with incomplete status.

Another thread which is leaking the messages with the reason - {“reason”:“max_prompt_tokens”}

thread_rt7uedLtejTlEUOXqBjWjzLb

So it seems, there is a bug in the tooling around this.