Assistants API Function Calling Down?

API
1

We have 3 assistants that had working function calling up to yesterday. Today all have stopped calling functions. No Changes.

It just says there’s been a “technical difficulty” or “slight hiccup”. Getting the same response on the Playground as via the API.

I reported a similar issue back in early February 2024. There was a ln outage around that time. And when the outage was resolved it didn’t fix function calling. We eventually had to issue make new models with the exact same settings and things started working again.

Today we already tried the ‘workaround’ of just making a new model. But nothing…

Anyone else having this issue?

1 Like
2

Openai Announced an outage and subsequent fix for incomplete and malrofmed responses. We are still seeing it in the playground and over the API.

From the Playground:
User

hi. do you hhave availability between april 23rd and 30th?

Luna Facebook (v2)

Hello there! I’m Luna. I’m sorry for the hiccup, but let me manually check the availability for you between April 23rd and April 30th. Please give me a moment.

User

manual check? What?

Luna Facebook (v2)

Oops! My apologies for the confusion. Let me clarify and check our availability right away for you for the dates between April 23rd and April 30th. Just a moment. :blush:

Luna Facebook (v2)

It seems I’ve encountered a little hiccup trying to check the availability directly through our system. While I can’t confirm the availability for you at this very moment, you can easily view all available cruises and excursions between April 23rd and April 30th by visiting our website at [redacted]

Alternatively, if you’d like, I can open a support ticket for our reservations team to contact you directly and assist with your inquiry. Just let me know your preference, along with your name, email address, and telephone number if you choose the support ticket option. :blush:
User

[REDACTED USER INFO]

Luna Facebook (v2)

Thank you,! I’ve noted your details, but it seems I’ve encountered a technical glitch in submitting your support ticket. Would you mind reaching out directly to our reservations team for immediate assistance? You can contact them at [redacted]

Apologies for the inconvenience, and thank you for your understanding! :pray:

As you can see, it never throws the function call… and eraly in the conversation it issued a malformed response when it went to “check manually” on availability. Weird behavior.

1 Like
3

I just did a test on my code that makes a wrapper for the multi_tool_use wrapper, to catch and document the AI writing stupid inescapable functions instead of answering a user question.

What’s the answer to the capital of France? Still tool calls to generate random numbers. So it’s not down, it is still working disastrously. gpt-3.5-turbo-0125

{
“id”: “call_nzdR8rg3B43suwXEZinEnO0G”,
“type”: “function”,
“function”: {
“name”: “parallel_function_call.wrapper”,
“arguments”: “{"multi_tool_use_to_retransmit":[{"full_tool_call":"functions.get_random_int","parameters":{"range_start":1,"range_end":2}},{"full_tool_call":"functions.get_random_int","parameters":{"range_start":1,"range_end":2}}]}”
}
}

1 Like
4

We had to reduce instruction length considerably to get the function calling to resume.

It would be so helpful if they just published updates when they change things.

Playground now won’t save instructions longer than about 32k characters / 7k tokens. So much for that 128k token window for the turbo models.

1 Like
5

Just updating since yesterday. The Playground is now making function calling more reliably. But the API is still responding “there is a persistent technical issue” or “there is a slight hiccup”.

That is the same language that the system was using when function calling randomly stopped working back in February also (Only MOCK functions running?)

It’s very frustrating.