Function Call: Bad Request

[rooislangwtf]

Hi,
I'm trying to recreate the python example of function calling in php (https://platform.openai.com/docs/guides/gpt/function-calling).
The example has 2 parts: part 1 does the call to ChatCompletion that includes the user message as well as the function(s) and part 2 that then gets a response from the suggested function which gets sent to ChatCompletion again to get a response based on the results from the function call. When I do part 2 I get an "Unsuccessful response. HTTP status code: 400 (Bad Request). on line 150 in ChatCompletions.php".

For the first successful call the request looks like this:

Tectalic\OpenAi\Models\ChatCompletions\CreateRequest Object
(
    [model] => gpt-3.5-turbo-0613
    [messages] => Array
        (
            [0] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => user
                    [content] => whats the weather like in boston
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

        )

    [functions] => Array
        (
            [0] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestFunctionsItem Object
                (
                    [name] => get_current_weather
                    [description] => Get the current weather in a given location
                    [parameters] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestFunctionsItemParameters Object
                        (
                            [value] => Array
                                (
                                    [type] => object
                                    [properties] => Array
                                        (
                                            [location] => Array
                                                (
                                                    [type] => string
                                                    [description] => The city and state, e.g. San Francisco, CA
                                                )

                                            [unit] => Array
                                                (
                                                    [type] => string
                                                    [enum] => Array
                                                        (
                                                            [0] => celsius
                                                            [1] => fahrenheit
                                                        )

                                                )

                                        )

                                    [required] => Array
                                        (
                                            [0] => location
                                        )

                                )

                            [ignoreMissing:protected] => 1
                            [exceptKeys:protected] => Array
                                (
                                )

                            [onlyKeys:protected] => Array
                                (
                                )

                        )

                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

        )

    [function_call] => auto
    [temperature] => 0.4
    [ignoreMissing:protected] => 1
    [exceptKeys:protected] => Array
        (
        )

    [onlyKeys:protected] => Array
        (
        )
)

And I get the expected response of:

Tectalic\OpenAi\Models\ChatCompletions\CreateResponse Object
(
    [id] => chatcmpl-7YWd5Y72513gzEDzPGpiKTAIX1cVx
    [object] => chat.completion
    [created] => 1688462935
    [model] => gpt-3.5-turbo-0613
    [choices] => Array
        (
            [0] => Tectalic\OpenAi\Models\ChatCompletions\CreateResponseChoicesItem Object
                (
                    [index] => 0
                    [message] => Tectalic\OpenAi\Models\ChatCompletions\CreateResponseChoicesItemMessage Object
                        (
                            [role] => assistant
                            [function_call] => Tectalic\OpenAi\Models\ChatCompletions\CreateResponseChoicesItemMessageFunctionCall Object
                                (
                                    [name] => get_current_weather
                                    [arguments] => {
  "location": "Boston, MA"
}
                                    [ignoreMissing:protected] => 1
                                    [exceptKeys:protected] => Array
                                        (
                                        )

                                    [onlyKeys:protected] => Array
                                        (
                                        )

                                )

                            [ignoreMissing:protected] => 1
                            [exceptKeys:protected] => Array
                                (
                                )

                            [onlyKeys:protected] => Array
                                (
                                )

                        )

                    [finish_reason] => function_call
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

        )

    [usage] => Tectalic\OpenAi\Models\ChatCompletions\CreateResponseUsage Object
        (
            [prompt_tokens] => 82
            [completion_tokens] => 18
            [total_tokens] => 100
            [ignoreMissing:protected] => 1
            [exceptKeys:protected] => Array
                (
                )

            [onlyKeys:protected] => Array
                (
                )

        )

    [ignoreMissing:protected] => 1
    [exceptKeys:protected] => Array
        (
        )

    [onlyKeys:protected] => Array
        (
        )

)

So for the second call I then make a response chain of:

Array
(
    [0] => Array
        (
            [role] => user
            [content] => whats the weather like in boston
        )

    [1] => Array
        (
            [role] => assistant
            [function_call] => Array
                (
                    [name] => get_current_weather
                    [arguments] => {"location":"Boston, MA"}
                )

            [content] => 
        )

    [2] => Array
        (
            [role] => function
            [name] => get_current_weather
            [content] => {"location":"Boston, MA","temperature":"72","unit":"fahrenheit","forecast":["sunny","windy"]}
        )

)

So the request looks like this which then causes the 400 error:

Tectalic\OpenAi\Models\ChatCompletions\CreateRequest Object
(
    [model] => gpt-3.5-turbo-0613
    [messages] => Array
        (
            [0] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => user
                    [content] => whats the weather like in boston
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

            [1] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => assistant
                    [function_call] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItemFunctionCall Object
                        (
                            [name] => get_current_weather
                            [arguments] => "{\"location\":\"Boston, MA\"}"
                            [ignoreMissing:protected] => 1
                            [exceptKeys:protected] => Array
                                (
                                )

                            [onlyKeys:protected] => Array
                                (
                                )

                        )

                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

            [2] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => function
                    [content] => {"location":"Boston, MA","temperature":"72","unit":"fahrenheit","forecast":["sunny","windy"]}
                    [name] => get_current_weather
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

        )

    [temperature] => 0.4
    [ignoreMissing:protected] => 1
    [exceptKeys:protected] => Array
        (
        )

    [onlyKeys:protected] => Array
        (
        )

)

I don't think the error is related to tokens etc. since I also did the python example and that works and if I remove the array that contains the "function_call" (array index 1) from the response chain I can get a response of "The current weather in Boston, MA is sunny and windy with a temperature of 72°F.".

Tectalic\OpenAi\Models\ChatCompletions\CreateRequest Object
(
    [model] => gpt-3.5-turbo-0613
    [messages] => Array
        (
            [0] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => user
                    [content] => whats the weather like in boston
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

            [1] => Tectalic\OpenAi\Models\ChatCompletions\CreateRequestMessagesItem Object
                (
                    [role] => function
                    [content] => {"location":"Boston, MA","temperature":"72","unit":"fahrenheit","forecast":["sunny","windy"]}
                    [name] => get_current_weather
                    [ignoreMissing:protected] => 1
                    [exceptKeys:protected] => Array
                        (
                        )

                    [onlyKeys:protected] => Array
                        (
                        )

                )

        )

    [temperature] => 0.4
    [ignoreMissing:protected] => 1
    [exceptKeys:protected] => Array
        (
        )

    [onlyKeys:protected] => Array
        (
        )

)

I've also hard coded a response chain in python and sent that to the chat completions endpoint and I also can get a response of "The current weather in Boston, MA is sunny and windy with a temperature of 72 degrees."

[
      {'role': 'user', 'content': "What's the weather like in Boston?"}, 
      {'role': 'assistant', 'content': None, "function_call": {"arguments": "{'location': 'Boston, MA'}", "name": "get_current_weather"}},
      {'role': 'function', 'name': 'get_current_weather', 'content': '{"location": "Boston, MA", "temperature": "72", "unit": null, "forecast": ["sunny", "windy"]}'}
]

So I can get it to work in php if I remove the message that contains the "function_call" response from the response chain, but I'm worried that would remove some needed context when I do real functions' responses. Just wanted to hear if there's a glaring error I'm making / if it's a bug?