basillangevin/instructor-laravel

Structured outputs for LLMs using Spatie Data objects.

v0.1.0 2025-02-11 21:53 UTC

README

Latest Version on Packagist GitHub Tests Action Status GitHub Code Style Action Status Total Downloads

Instructor Laravel turns LLM (Large Language Model) responses into spatie/laravel-data objects, helping you easily integrate LLMs into your applications.

Installation

You can install the package via composer:

composer require basillangevin/instructor-laravel

You can publish the config file with:

php artisan vendor:publish --tag="instructor-laravel-config"

These are the contents of the published config file:

use EchoLabs\Prism\Enums\Provider;

// config for BasilLangevin/InstructorLaravel
return [
    /*
    |--------------------------------------------------------------------------
    | Default LLM Provider
    |--------------------------------------------------------------------------
    |
    | This value is the default LLM provider that this package will use to
    | generate a structured response that it will transform into a Data
    | object. You may also set LLM providers on a per-request basis.
    */
    'provider' => Provider::OpenAI,

    /*
    |--------------------------------------------------------------------------
    | Default LLM Model
    |--------------------------------------------------------------------------
    |
    | This value is the default LLM model that this package will use
    | for the LLM provider when generating a structured response.
    | You may also set the model each time you make a request.
    */
    'model' => 'gpt-4o',
];

Introduction

Instructor Laravel is built on top of echolabsdev/prism: a fantastic package that provides a unified interface to work with various LLM providers.

Instructor Laravel extends the Structured Outputs feature of Prism, allowing you to use spatie/laravel-data objects as your data schema.

The package modifies two methods from Prism: withSchema and generate. All other Prism methods are available and work as expected.

Basic Usage

Because this package transforms LLM responses into spatie/laravel-data objects, we'll start by defining a Data object.

use Spatie\LaravelData\Attributes\Min;
use Spatie\LaravelData\Data;

class BirdData extends Data
{
    public function __construct(
        public string $species,

        /** The average wingspan of the bird in centimeters. */
        public int $wingspan,

        #[In(['forest', 'prarie', 'wetland'])]
        public string $habitat,
    ) {}
}

Then, we'll build our LLM request by:

  • starting with the Instructor facade,
  • passing our Data object class to the withSchema method,
  • using standard Prism structured output methods to configure the request, then
  • calling the generate method.
use BasilLangevin\InstructorLaravel\Facades\Instructor;
use EchoLabs\Prism\Enums\Provider;

$bird = Instructor::make()
    ->withSchema(BirdData::class)
    ->using(Provider::OpenAI, 'gpt-4o')
    ->withPrompt('Tell me about a bird found on the West Coast of Canada.')
    ->generate();

The generate method will return a BirdData object from the LLM response.

BirdData {
  +species: "Western Bluebird"
  +wingspan: 34
  +habitat: "forest"
}

Collections of Data objects

Instructor Laravel also supports transforming responses that contain collections of Data objects.

$birds = Instructor::make()
    ->withCollectionSchema(BirdData::class)
    ...

The generate method will return a Collection`` of BirdData` objects.

You may pass a custom Collection class name as the second argument of withCollectionSchema to use a custom collection class:

$birds = Instructor::make()
    ->withCollectionSchema(BirdData::class, BirdCollection::class)
    ...

Validation and retries

Instructor Laravel will automatically validate the LLM response against the Data object's rules.

If the LLM response doesn't match the expected structure, Instructor Laravel will call the LLM again with any errors it encountered, helping the LLM generate a response that matches the expected structure.

By default, the request will be retried up to 3 times. You can change this by calling withRetries(...) or withoutRetries():

$bird = Instructor::withSchema(BirdData::class)
    ->withRetries(5)
    ...

The withRetries method accepts the same arguments as Laravel's retry helper (excluding the $callback argument). The number of retries will be added to the initial request, so if you pass 3 to withRetries, the request will be made up to 4 times in total (1 initial attempt + 3 retries).

Improving LLM response reliability

To improve the LLM's understanding of the expected response, Laravel Instructor transforms the Data class into a JSON Schema and passes it to the LLM.

Under the hood, schemas are generated using basillangevin/laravel-json-data-schema. This package adds additional information to the generated schema by transforming:

All of this information is available to the LLM from the very first request, so it can be helpful to annotate your Data classes more information as needed.

Testing

Each PHP file in this package has a co-located Pest test file named {FileName}Test.php.

This package achieves 100% test coverage, 100% mutation coverage, and 100% PHPStan coverage coverage at level 10.

The following commands can be used to test the package:

# Run the standard test suite
./vendor/bin/pest --parallel

# Run the test suite and generate a coverage report
./vendor/bin/pest --coverage --parallel

# Run mutation tests
./vendor/bin/pest --mutate --parallel --covered-only

# Run PHPStan at level 10
./vendor/bin/phpstan analyse

Changelog

Please see CHANGELOG for more information on what has changed recently.

Contributing

Please see CONTRIBUTING for details.

Security Vulnerabilities

Please review our security policy on how to report security vulnerabilities.

Credits

License

The MIT License (MIT). Please see License File for more information.