Please Correct Me. Assistant API

API
1

i am linking my Assistant API to my Bubble APPS.
CURL call only

My step after reading Offical Doc :
( pre set Asst ID )

  1. User Input

  2. Create Thread and Run

  3. [this part confuse me] every 0.5 sec check Run Status, If complete goto step 4.

  4. List Message and display.

Why i cant use Assistant API like GPT 3.5 API? just sent input and wait for response?

from Official Doc:
By default, a Run goes into the queued state. You can periodically retrieve the Run to check on its status to see if it has moved to completed.

so i need to create a workflow that looping the check run status to get the response?

Please correct me if i am wrong. Thanks in advance.

2

Hi and welcome to the community.
Yes you can list the messages added to the thread once the run completes.
I think you already referred to the “How assistants work” part of the documentation.
Here it is a bit more clear:
https://platform.openai.com/docs/assistants/overview

You can check the run steps for a bit more transparency:
https://platform.openai.com/docs/api-reference/runs/listRunSteps

2 Likes
3

Based on my Assistants I would recommend a ping >1 second. I have settled for 2 seconds.

4

Yep. We’re just doing something like this:

bool isComplete = false;

while (!isComplete)
{
    // No point checking for completion right after starting the run anyway.
    await Task.Delay(TimeSpan.FromSeconds(5), cancellationToken);

    var threadRun = await _openAiHttpClient.GetThreadRunAsync(threadRunResult.ThreadId, threadRunResult.Id, cancellationToken);
    if (threadRun.Status == "failed")
    {
        throw new InvalidOperationException($"Error: {threadRun}");
    }

    isComplete = threadRun.Status == "completed";
}

Being able to listen over an open TCP connection is coming, apparently. It’s in their interest as well since polling like this is likely more overhead for them.

5

Yes, you should create a loop. My recommendation is that you check the status every second and, after a testing period, observe the average in your responses to make the necessary adjustments. In my case, for text, I have an average response time of 3 seconds and 30 seconds when using a code interpreter. Initially, you will have to sacrifice workload units :frowning: