OpenAI Function Calling
When making an API call to models gpt-3.5-turbo-0613 & gpt-4-0613, users can describe a function. The model generates a JSON output containing arguments.
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
To be clear, the chat completion API does not call any function but the model does generate the JSON which can be used to call a function from your code.
Consider the following working API call:
An API call is submitted to https://api.openai.com/v1/chat/completions as seen below, with the OPEN_API_KEY defined in the header.
Below is the JSON document sent to the model. The aim of this call is to generate a JSON file which can be used to send to an API which sends emails.
You can see the name is send_email. And three parameters are defined, to_address , subject and body , which is the email body. The chatbot user request is: Send Cobus from humanfirst ai an email asking for the monthly report?
{
"model": "gpt-3.5-turbo-0613",
"messages": [
{"role": "user", "content": "Send Cobus from humanfirst ai an email asking for the monthly report?"}
],
"functions": [
{
"name": "send_email",
"description": "Please send an email.",
"parameters": {
"type": "object",
"properties": {
"to_address": {
"type": "string",
"description": "To address for email"
},
"subject": {
"type": "string",
"description": "subject of the email"
},
"body": {
"type": "string",
"description": "Body of the email"
}
}
}
}
]
}Below is the JSON generated by the model, in this case the model gpt-3.5-turbo-0613 is used for the completion.
{
"id": "chatcmpl-7R3k9pN6lLXCmNMiLNoaotNAg86Qg",
"object": "chat.completion",
"created": 1686683601,
"model": "gpt-3.5-turbo-0613",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": null,
"function_call": {
"name": "send_email",
"arguments": "{\n \"to_address\": \"cobus@humanfirst.ai\",\n \"subject\": \"Request for Monthly Report\",\n \"body\": \"Dear Cobus,\\n\\nI hope this email finds you well. I would like to kindly request the monthly report for the current month. Could you please provide me with the report by the end of the week?\\n\\nThank you in advance for your assistance!\\n\\nBest regards,\\n[Your Name]\"\n}"
}
},
"finish_reason": "function_call"
}
],
"usage": {
"prompt_tokens": 86,
"completion_tokens": 99,
"total_tokens": 185
}
}This is a big leap in the right direction, with the Large Language Model not only structuring output into natural conversational language.
But structuring output in a format to be consumed by another system, as apposed to a human.
OpenAI Function Calling structures output for machine consumption in the form of an API, as apposed to human consumption in the form of unstructured natural language.
There are a few considerations, however:
- Programmatically the chatbot/Conversational UI will have to know that the output to the LLM must be in JSON format. Hence there will have to be some classification of sorts to detect output type.
- A predefined template will have to exist, for input to the completion LLM. As seen below, the JSON template guides the LLM on how to populate the values.
- More importantly, as seen below in the image, the parameters which should be populated, must be well defined. Failure results in the following error:
“We could not parse the JSON body of your request. (HINT: This likely means you aren’t using your HTTP library correctly. The OpenAI API expects a JSON payload, but what was sent was not valid JSON. If you have trouble figuring out how to fix this, please contact us through our help center at help.openai.com.)” - And again, this is a big step in the right direction, but there is still an element of rigidity in assigning values to the JSON parameters.
Below is another use-case to consider. A travel booking example, with the parameters defined as:
destination, departure, number_people and travel_model.
The user utterance is defined as: I need to book a trip from Bonn to Amsterdam for my wife, mother and by two sons and daughter. I will also be joining them. The airline must fly direct.
{
"model": "gpt-3.5-turbo-0613",
"messages": [
{"role": "user", "content": "I need to book a trip from Bonn to Amsterdam for my wife, mother and by two sons and daughter. I will also be joining them. The airline must fly direct."}
],
"functions": [
{
"name": "book_travel",
"description": "Book travel",
"parameters": {
"type": "object",
"properties": {
"destination": {
"type": "string",
"description": "Your travel destination."
},
"departute": {
"type": "string",
"description": "From where are you traveling"
},
"number_people": {
"type": "string",
"description": "How many people are traveling"
},
"travel_mode": {
"type": "string",
"description": "What mode of travel will it be."
}
}
}
}
]
}And the output below:
{
"id": "chatcmpl-7R3vsPC6JndweAQZXvCIvSWnaXDMP",
"object": "chat.completion",
"created": 1686684328,
"model": "gpt-3.5-turbo-0613",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": null,
"function_call": {
"name": "book_travel",
"arguments": "{\n \"destination\": \"Amsterdam\",\n \"departute\": \"Bonn\",\n \"number_people\": \"6\",\n \"travel_mode\": \"airline\"\n}"
}
},
"finish_reason": "function_call"
}
],
"usage": {
"prompt_tokens": 122,
"completion_tokens": 42,
"total_tokens": 164
}
}The primary use-cases of this function will be:
- Creating chatbots to answer questions by calling external APIs.
- Convert natural language into structured JSON data.
- Extract structured data from text.
⭐️ Please follow me on LinkedIn for updates on Conversational AI ⭐️
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
