Loading...
Skip to main content

Prompting in AI.JSX

Prerequisites​

Before jumping into this guide on prompting, there are a couple of other guides we suggest you read first:

The Power of Chat-Based Models​

A core part of AI apps is prompting the model and getting a response. To do that in AI.JSX, start with ChatCompletion:

function App() {
  return (
    <ChatCompletion>
      <UserMessage>Tell me a story about ducks.</UserMessage>
    </ChatCompletion>
  );
}

ChatCompletion is preferred to Completion because all the most powerful models are chat-based, and it's best to start with the most powerful models.

To configure the output of ChatCompletion, use ModelProps. This allows you to do things like making the model more creative or precise, telling the model how long a response you want back, etc. Combined with the natural language of your prompt, this is how you control the model's output.

The Problem with Non-Chat Models​

Chat models have stronger "instructability". This just means they follow your instructions better.

Conversely, non-chat completion models are just trained to predict the next token. With a non-chat model, you might see something like this:

Input:
  what is the capital of North Dakota?

Output:
  what is the capital of South Dakota?
  what is the capital of Indiana?
  what is the capital of New York?

Instead of giving us "Bismarck", the actual capital of North Dakota, the model is predicting that a question about one state is often followed by questions about other states.

Constrained Output​

Sometimes it's important we get the AI model to respond in a particular way. We want the model to respond with the correct information and present the information in a specific way. For example, we might want the response to be formatted as JSON or YAML.

Doing this reliably can be hard which is why AI.JSX provides the "constrained-output" module to help with this type of formatting. Under the hood, the module uses a combination of prompting, validating the output, and retries if the validation fails. See ai-jsx/batteries/constrained-output for more information.

Example: Returning JSON and YAML​

function App() {
  return (
    <>
      JSON generation example:{'\n'}
      <JsonChatCompletion>
        <UserMessage>
          Create a random object describing an imaginary person that has a "name", "gender", and "age".
        </UserMessage>
      </JsonChatCompletion>
      {'\n\n'}
      YAML generation example:{'\n'}
      <YamlChatCompletion>
        <UserMessage>
          Create a random object describing an imaginary person that has a "name", "gender", and "age".
        </UserMessage>
      </YamlChatCompletion>
    </>
  );
}

Custom Schema Enforcement​

You can also provide your own schema and then have AI.JSX enforce it in the output:

// We use `zod` library to create and enforce the schema
import z from 'zod';

const FamilyTree: z.Schema = z.array(
  z.object({
    name: z.string(),
    children: z.lazy(() => FamilyTree).optional(),
  })
);

function App() {
  return (
    <JsonChatCompletion schema={FamilyTree}>
      <UserMessage>Create a nested family tree with names and ages. It should include a total of 5 people</UserMessage>
    </JsonChatCompletion>
  );
}