Introduction

What is Web Intents for LLMs?

With the rise of chatbots and other conversational interfaces, it's important to have a standard way to define the capabilities of a web service. This is where Web Intents for LLMs comes in. It's a protocol that allows web services to define their capabilities in a standard way so that chatbots and other conversational interfaces can easily interact with them.

I've thought about using the term "Web Intents". Although it shares some commonalities with Web Intents, it is not the same thing as Google's proposed Web Intents API. So, to avoid any confusion, We've added "for LLMs" to the name.

Example

Web Intents for LLMs

The LLMs recognize the intent and then embed the interface in the chat. The user can then interact with the interface to get the information they need. Here is the web page that is embedded in the chat:

Breaking down the example

The intents.json

LLMs accept function calls, so we have to define the function that the LLMs will call. Here's the definition modified form OpenAI's function definition in JSON format:

{
    "name": "is_streamer_live_on_twitch",
    "description": "Check if a streamer is live on Twitch by their login name.",
    "parameters": {
        "type": "object",
        "properties": {
            "login_name": {
                "type": "string",
                "description": "The streamer's login name."
            }
        },
        "required": [
            "login_name"
        ]
    },
    "async": true,
    "link": "/?name={login_name}",
    "example": "/?name=ninja",
    "returns": {
        "type": "object",
        "properties": {
            "response": {
                "type": "string",
                "description": "The information of the streamer."
            }
        }
    }
}

With this definition, LLMs based chatbots can call the function is_streamer_live_on_twitch by extracting the function call related information from JSON and pass it to the completion APIs.

You can check OpenAI's completion API here (opens in a new tab). The difference between OpenAI's completion API and this is that we add the following fields:

  1. link: The link to the web service that will be called.
  2. example: An example of the link with the parameters filled in.
  3. returns: The return value of the function.

You can also visit the json file at https://spellboard-twitch-intent.vercel.app/intents.json (opens in a new tab)

Demo Videos

 

You can find more videos at here (opens in a new tab)

How to make it run on SpellBoard

  1. create a intents.json file and place it in the root of your web service.
  2. make your site embedable by disable the iframe protection.
    1. X-Frame-Options: set the header of your site to ALLOW-FROM https://spellboard.getmegaportal.com/ or disable it.
    2. Content-Security-Policy: set the header of your site to frame-ancestors 'self' https://spellboard.getmegaportal.com/ or disable it.
    3. Make your cookie SameSite=None; Secure if you are using cookies.
    4. Any other security measures that you have in place that might prevent your site from being embedded in an iframe.
  3. Visit https://spellboard.getmegaportal.com/?appUrl=${encodeURIComponent('https://your-site.com')} to test your site.

Try the demo yourself

You can try the demo yourself by visiting here (opens in a new tab) and then clicking the "Chat" button.

Demo Source Code

You can find the source code for the demo at here (opens in a new tab)

Triple-Win Strategy for Enhancing User Experiences with LLM-Based Platforms

We think here's how it may work from a triple-win perspective:

  1. Users care about the outcome and enriched user experiences without needing to learn how to use the app first. So, cut to the chase.

  2. Third-party app providers want their data protected and not leaked to LLM providers. Most importantly, it provides a way to convert traffic to their own sites, similar to what they expect from a search engine, rather than being replaced by a feature of an LLM-based app.

  3. LLM-based platforms require little effort to support this and can monetize by replicating traditional methods of recommending and advertising apps.

Versioning