Tool Calls Stopped Firing Automatically

API
1

Has there been any recent changes to the API. Just over the last day or so my tool calls stopped firing automatically. I have a chatbot that can perform various functions and on the message that they are about to perform a function, it says something along the lines of “Just one moment while I …” the tool call would immediately be fired. But just recently I now have to respond to that message with an ok or go ahead for it to fire off. What could be the cause of that?

2

An AI that doesn’t actually have the tool will produce that kind of response, I have seen. Language to you, but it has no method to invoke - or it would have.

OpenAI just yesterday introduced a “forced function calling” feature, so this may have taken back-end manipulation of chat completions causing symptoms.

I can report that gpt-3.5-turbo-0125 is just as dumb as ever on a script from Feb 2.

  • I inject a tool wrapper to override the multi_tool_use shim of parallel tool calls for capturing what it would emit to that tool.
  • I ask it about the capital of France and Germany
  • It emits the tokens of calling tools instead of writing a response to me.
  • It has no use for functions, but can’t escape and has to put something in just functions that exist - and we get random number functions to answer about capital cities.

{
“id”: “call_LZlJoohXklaJ1058668IjJkp”,
“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}}]}”
}
}

So the API and interface for that model, which can run on assistants, is unaffected. You can see how your current specification meshes with the API reference and that new parameter.

2 Likes