Skip to main content

​
Overview

Sometimes property level validation is not enough - you may want to check values of multiple properties and based on the combination of them decide to accept or reject the response. Or the assertions provided by Symfony may not be enough for your use case. In such case you can easily add custom validation code to your response model by:
  • using ValidationMixin
  • and defining validation logic in validate() method.
In this example LLM should be able to correct typo in the message (graduation year we provided is 1010 instead of 2010) and respond with correct graduation year.

​
Example

<?php
require 'examples/boot.php';

use Cognesy\Instructor\StructuredOutput;
use Cognesy\Instructor\StructuredOutputRuntime;
use Cognesy\Instructor\Validation\Traits\ValidationMixin;
use Cognesy\Instructor\Validation\ValidationResult;
use Cognesy\Polyglot\Inference\LLMProvider;

class UserDetails
{
    use ValidationMixin;

    public string $name;
    public int $birthYear;
    public int $graduationYear;

    public function validate() : ValidationResult {
        if ($this->graduationYear > $this->birthYear) {
            return ValidationResult::valid();
        }
        return ValidationResult::fieldError(
            field: 'graduationYear',
            value: $this->graduationYear,
            message: "Graduation year has to be bigger than birth year."
        );
    }
}

$runtime = StructuredOutputRuntime::fromProvider(LLMProvider::using('openai'))
    ->withMaxRetries(2)
    ->wiretap(fn($e) => $e->print());

$user = (new StructuredOutput($runtime))
    ->withResponseClass(UserDetails::class)
    ->with(
        messages: [['role' => 'user', 'content' => 'Jason was born in 2000 and graduated in 18.']],
        model: 'gpt-4o-mini',
    )->get();


dump($user);

assert($user->graduationYear === 2018);
?>