OpenAI’s “Operate Calling” is perhaps probably the most groundbreaking but underneath appreciated characteristic launched by any software program firm… ever.
Capabilities help you flip unstructured information into structured information. This may not sound all that groundbreaking however when you think about that 90% of information processing and information entry jobs worldwide exist for this precise purpose, it’s fairly a revolutionary characteristic that went considerably unnoticed.
Have you ever ever discovered your self begging GPT (3.5 or 4) to spit out the reply you need and completely nothing else? No “Positive, right here is your…” or every other ineffective fluff surrounding the core reply. GPT Capabilities are the answer you’ve been searching for.
How are Capabilities meant to work?
OpenAI’s docs on perform calling are extraordinarily restricted. You’ll end up digging via their developer discussion board for examples of find out how to use them. I dug across the discussion board for you and have many instance developing.
Right here’s one of many solely examples you’ll be capable to discover of their docs:
features = [
{
"name": "get_current_weather",
"description": "Get the current weather in a given location",
"parameters": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "The city and state, e.g. San Francisco, CA",
},
"unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
},
"required": ["location"],
},
}
]
A perform definition is a inflexible JSON format that defines a perform title, description and parameters. On this case, the perform is supposed to get the present climate. Clearly GPT isn’t in a position to name this particular API (because it doesn’t exist) however utilizing this structured response you’d be capable to join the actual API hypothetically.
At a excessive stage nevertheless, features present two layers of inference:
Choosing the perform itself:
Chances are you’ll discover that features are handed into the OpenAI API name as an array. The rationale you present a reputation and outline to every perform are so GPT can determine which to make use of based mostly on a given immediate. Offering a number of features in your API name is like giving GPT a Swiss military knife and asking it to chop a chunk of wooden in half. It is aware of that despite the fact that it has a pair of pliers, scissors and a knife, it ought to use the noticed!
Operate definitions contribute in the direction of your token depend. Passing in a whole bunch of features wouldn’t solely take up the vast majority of your token restrict but additionally end in a drop in response high quality. I usually don’t even use this characteristic and solely go in 1 perform that I power it to make use of. It is extremely good to have in sure use circumstances nevertheless.
Choosing the parameter values based mostly on a immediate:
That is the actual magic in my view. GPT having the ability to select the software in it’s software equipment is wonderful and undoubtedly the main target of their characteristic announcement however I believe this is applicable to extra use circumstances.
You possibly can think about a perform like handing GPT a type to fill out. It makes use of its reasoning, the context of the state of affairs and area names/descriptions to determine the way it will fill out every area. Designing the shape and the extra info you go in is the place you may get inventive.
Probably the most frequent issues I take advantage of features for to extract particular values from a big chunk of textual content. The sender’s handle from an e-mail, a founders title from a weblog submit, a telephone quantity from a touchdown web page.
I prefer to think about I’m looking for a needle in a haystack besides the LLM burns the haystack, leaving nothing however the needle(s).
Use case: Processing hundreds of contest submissions
I constructed an automation that iterated over hundreds of contest submissions. Earlier than storing these in a Google sheet I wished to extract the e-mail related to the submission. Heres the perform name I used for extracting their e-mail.
{
"title":"update_email",
"description":"Updates e-mail based mostly on the content material of their submission.",
"parameters":{
"kind":"object",
"properties":{
"e-mail":{
"kind":"string",
"description":"The e-mail supplied within the submission"
}
},
"required":[
"email"
]
}
}
Assigning unstructured information a rating based mostly on dynamic, pure language standards is a superb use case for features. You possibly can rating feedback throughout sentiment evaluation, essays based mostly on a customized grading rubric, a mortgage utility for threat based mostly on key elements. A current use case I utilized scoring to was scoring of gross sales leads from 0–100 based mostly on their viability.
Use Case: Scoring Gross sales leads
We had a whole bunch of potential leads in a single google sheet a number of months in the past that we wished to sort out from most to least vital. Every lead contained data like firm measurement, contact title, place, trade and so forth.
Utilizing the next perform we scored every lead from 0–100 based mostly on our wants after which sorted them from greatest to worst.
{
"title":"update_sales_lead_value_score",
"description":"Updates the rating of a gross sales lead and gives a justification",
"parameters":{
"kind":"object",
"properties":{
"sales_lead_value_score":{
"kind":"quantity",
"description":"An integer worth starting from 0 to 100 that represents the standard of a gross sales lead based mostly on these standards. 100 is an ideal lead, 0 is horrible. Perfect Lead Standards:n- Medium sized firms (300-500 staff is the most effective vary)n- Corporations in main useful resource heavy industries are greatest, ex. manufacturing, agriculture, and so forth. (that is crucial standards)n- The upper up the contact place, the higher. VP or Government stage is most well-liked."
},
"score_justification":{
"kind":"string",
"description":"A transparent and conscise justification for the rating supplied based mostly on the customized standards"
}
}
},
"required":[
"sales_lead_value_score",
"score_justification"
]
}
Outline customized buckets and have GPT thoughtfully think about each bit of information you give it and place it within the appropriate bucket. This can be utilized for labelling duties like choosing the class of youtube movies or for discrete scoring duties like assigning letter grades to homework assignments.
Use Case: Labelling information articles.
A quite common first step in information processing workflows is separating incoming information into completely different streams. A current automation I constructed did precisely this with information articles scraped from the net. I wished to kind them based mostly on the subject of the article and embrace a justification for the choice as soon as once more. Right here’s the perform I used:
{
"title":"categorize",
"description":"Categorize the enter information into consumer outlined buckets.",
"parameters":{
"kind":"object",
"properties":{
"class":{
"kind":"string",
"enum":[
"US Politics",
"Pandemic",
"Economy",
"Pop culture",
"Other"
],
"description":"US Politics: Associated to US politics or US politicians, Pandemic: Associated to the Coronavirus Pandemix, Financial system: Associated to the financial system of a particular nation or the world. , Popular culture: Associated to popular culture, movie star media or leisure., Different: Does not slot in any of the outlined classes. "
},
"justification":{
"kind":"string",
"description":"A brief justification explaining why the enter information was categorized into the chosen class."
}
},
"required":[
"category",
"justification"
]
}
}
Usually occasions when processing information, I give GPT many attainable choices and wish it to pick the most effective one based mostly on my wants. I solely need the worth it chosen, no surrounding fluff or further ideas. Capabilities are good for this.
Use Case: Discovering the “most attention-grabbing AI information story” from hacker information
I wrote one other medium article right here about how I automated my total Twitter account with GPT. A part of that course of includes choosing probably the most related posts from the entrance pages of hacker information. This submit choice step leverages features!
To summarize the features portion of the use case, we might scrape the primary n pages of hacker information and ask GPT to pick the submit most related to “AI information or tech information”. GPT would return solely the headline and the hyperlink chosen through features in order that I might go on to scrape that web site and generate a tweet from it.
I might go within the consumer outlined question as a part of the message and use the next perform definition:
{
"title":"find_best_post",
"description":"Decide the most effective submit that almost all intently displays the question.",
"parameters":{
"kind":"object",
"properties":{
"best_post_title":{
"kind":"string",
"description":"The title of the submit that almost all intently displays the question, acknowledged precisely because it seems within the checklist of titles."
}
},
"required":[
"best_post_title"
]
}
}
Filtering is a subset of categorization the place you categorize gadgets as both true or false based mostly on a pure language situation. A situation like “is Spanish” will be capable to filter out all Spanish feedback, articles and so forth. utilizing a easy perform and conditional assertion instantly after.
Use Case: Filtering contest submission
The identical automation that I discussed within the “Information Extraction” part used ai-powered-filtering to weed out contest submissions that didn’t meet the deal-breaking standards. Issues like “should use typescript” have been completely necessary for the coding contest at hand. We used features to filter out submissions and trim down the overall set being processed by 90%. Right here is the perform definition we used.
{
"title":"apply_condition",
"description":"Used to determine whether or not the enter meets the consumer supplied situation.",
"parameters":{
"kind":"object",
"properties":{
"determination":{
"kind":"string",
"enum":[
"True",
"False"
],
"description":"True if the enter meets this situation 'Does submission meet the ALL these necessities (makes use of typescript, makes use of tailwindcss, useful demo)', False in any other case."
}
},
"required":[
"decision"
]
}
}
In case you’re curious why I really like features a lot or what I’ve constructed with them it is best to try AgentHub!
AgentHub is the Y Combinator-backed startup I co-founded that permit’s you automate any repetitive or complicated workflow with AI through a easy drag and drop no-code platform.
“Think about Zapier however AI-first and on crack.” — Me
Automations are constructed with particular person nodes referred to as “Operators” which might be linked collectively to create energy AI pipelines. We now have a list of AI powered operators that leverage features underneath the hood.
Try these templates to see examples of perform use-cases on AgentHub: Scoring, Categorization, Possibility-Choice,
If you wish to begin constructing AgentHub is dwell and able to use! We’re very energetic in our discord neighborhood and are joyful that will help you construct your automations if wanted.
Be at liberty to comply with the official AgentHub twitter for updates and myself for AI-related content material.
