Here there are a number of steps. The above instance makes use of Assistants API however the identical ideas apply to the chat completion endpoint as effectively.
Scenario — We need to have the ability to work together with the agent or ai assitant and ask it to research how the information articles for a selected firm are affecting its inventory value.
Step 0 — Identify the API companies you need to use. Here we have now 2 APIs we want. we want a Stock Price API service that takes the inventory image and returns costs for the required dates. We additionally want a second API service that takes within the identify of the corporate and searches for articles between a begin date and finish date. Once recognized we have now to research the construction through which these API requests have to be made. We take that construction and put it in a JSON format within the Assistants or Chat completions instruments part we name this construction the operate name. This step is essential as a result of it offers the assistant a blueprint on what format to construction its output. The schema is openapi schema. discover an instance under.
{
"identify": "get_weather",
"description": "Determine climate in my location",
"parameters": {
"kind": "object",
"properties": {
"location": {
"kind": "string",
"description": "The metropolis and state e.g. San Francisco, CA"
},
"unit": {
"kind": "string",
"enum": [
"c",
"f"
]
}
},
"required": [
"location"
]
}
}Step 1 — This is the the place we’re interacting with our AI assistant. We are requesting that the assistant analyzes articles and inventory value for Amazon and return its evaluation. Here the calls are despatched to the OpenAI api endpoint (both assistants api or chat completion)
Step 2 & 3 – The OpenAI mannequin then goes by way of the chat and determines the precise approach to output the two operate calls. Here the mannequin determines the Stock ticker identify. It additionally determines the beginning dates and finish dates primarily based on our pure language chat. This is the place the true heft of utilizing LLMs as backends comes by way of, No matter what the person interplay is the LLM can act as a layer and construction the output. The operate name structured outputs are nested within the instruments part.
Step 4- The mannequin then sends a ready state to the appliance (that is the appliance that the person is utilizing , in my instance i’ve flutterflow as backend and frontend).
Step 5 & 6- The person utility then sends these API calls to the API companies. The Application might want to deal with authentication to get the API to return legitimate knowledge as outputs.
Step 7 — The API responses or the standing of the responses are then despatched to the ready state run that was despatched as a part of the LLMs name.
Step 8 — The Chat Completion or Assistant API then returns the following message as a part of the AI assistant interplay. The LLM can both analyze the outcomes primarily based on the returned knowledge and use its comprehension capability to reply the query.
Biggest Takeaways
The largest takeaways from this are to grasp the totally different calls concerned so we are able to setup the API coordination on our utility finish.
Parallel Function calling
There are instances when there are 2 or extra capabilities they get referred to as on the similar time. See under my textual content triggered the circumstances for each the calls.
Forced Function calling and to skip operate calling.
You may also drive a selected operate to be referred to as by specifying the software selection parameter. Likewise you might need to not invoke a operate in that case ensure that software selection is none. see under the main points about that.
Link — Controls which (if any) operate is named by the mannequin. none means the mannequin is not going to name a operate and as a substitute generates a message. auto means the mannequin can decide between producing a message or calling a operate. Specifying a selected operate by way of {"kind: "operate", "operate": {"identify": "my_function"}} forces the mannequin to name that operate.
Essential hyperlinks to learn up on
OpenAI’s cookbook regarding the operate calling examples right here.
Flutterflow overview Demo video right here.
OpenAI
How did I construct my demo
I utilized all of the above ideas into my construct for my app. I constructed out the identical construction and deployed one thing like this.
