Structured Output

The LabsLLM allows you to define structured output schemas for your prompts, ensuring that the model's response follows a specific format.

Defining Output Schema

You can define an output schema using the same parameter types used for function definitions:

use LabsLLM\Parameters\ObjectParameter;
use LabsLLM\Parameters\StringParameter;
use LabsLLM\Parameters\NumberParameter;
use LabsLLM\Parameters\ArrayParameter;
use LabsLLM\Parameters\BooleanParameter;

// Define a simple structured output for a product review
$structureObject = new ObjectParameter('review', 'Product review data', [
    new StringParameter('title', 'Title of the review'),
    new NumberParameter('rating', 'Rating from 1 to 5'),
    new StringParameter('comment', 'Detailed review comment'),
    new ArrayParameter('tags', 'Review tags', 
        new StringParameter('tag', 'A single tag')
    ),
    new BooleanParameter('recommended', 'Whether the product is recommended')
], ['title', 'rating', 'comment']);

// Use the schema in your prompt
$execute = LabsLLM::text()
    ->using(new OpenAI('your-api-key', 'gpt-4'))
    ->withOutputSchema($structureObject)
    ->executePrompt('Write a review for the latest iPhone model');

$response = $execute->getResponseData();

Response Format

When using structured output, you can get the structured response using getStructureResponse():

// Get the structured response object
$reviewData = $response->getStructureResponse();

echo $reviewData->title;     // "Amazing iPhone Experience"
echo $reviewData->rating;    // 5
echo $reviewData->comment;   // "The new iPhone exceeded all my expectations..."

// Accessing array properties
foreach ($reviewData->tags as $tag) {
    echo $tag->tag;         // "battery", "camera", etc.
}

// Accessing boolean properties
if ($reviewData->recommended) {
    echo "This product is recommended!";
}

Best Practices

1. Keep schemas simple: Start with basic structures and add complexity as needed
2. Use clear field names: Choose descriptive names that clearly indicate the purpose of each field
3. Validate responses: Always validate the JSON response matches your expected structure
4. Handle errors: Be prepared to handle cases where the model might not follow the schema exactly

Important: While the model will try to follow your schema, it's important to validate the response as the model might occasionally deviate from the exact structure you defined.