JSON Schema Mode

JSON Schema mode takes JSON generation a step further by validating the response against a predefined schema. This ensures the response has exactly the structure your application expects.

Defining and Using a JSON Schema

<?php
use Cognesy\Polyglot\Inference\Inference;
use Cognesy\Polyglot\Inference\Enums\OutputMode;

$inference = new Inference()->using('openai');  // Currently best supported by OpenAI

// Define a schema for a weather report
$schema = [
    'type' => 'object',
    'properties' => [
        'location' => [
            'type' => 'string',
            'description' => 'The city and country'
        ],
        'current_temperature' => [
            'type' => 'number',
            'description' => 'Current temperature in Celsius'
        ],
        'conditions' => [
            'type' => 'string',
            'description' => 'Current weather conditions (e.g., sunny, rainy)'
        ],
        'forecast' => [
            'type' => 'array',
            'items' => [
                'type' => 'object',
                'properties' => [
                    'day' => [
                        'type' => 'string',
                        'description' => 'Day of the week'
                    ],
                    'temperature_high' => [
                        'type' => 'number',
                        'description' => 'Expected high temperature in Celsius'
                    ],
                    'temperature_low' => [
                        'type' => 'number',
                        'description' => 'Expected low temperature in Celsius'
                    ],
                    'conditions' => [
                        'type' => 'string',
                        'description' => 'Expected weather conditions'
                    ]
                ],
                'required' => ['day', 'temperature_high', 'temperature_low', 'conditions']
            ],
            'description' => 'Three-day weather forecast'
        ]
    ],
    'required' => ['location', 'current_temperature', 'conditions', 'forecast']
];

// Request a weather report
$response = $inference->with(
    messages: 'Provide a weather report for Paris, France.',
    responseFormat: [
        'type' => 'json_schema',
        'json_schema' => [
            'name' => 'weather_report',
            'schema' => $schema,
            'strict' => true,
        ],
    ],
    mode: OutputMode::JsonSchema
)->asJsonData();

// The response will match the schema's structure exactly
echo "Weather in {$response['location']}:\n";
echo "Currently {$response['conditions']} and {$response['current_temperature']}°C\n\n";

echo "Forecast:\n";
foreach ($response['forecast'] as $day) {
    echo "{$day['day']}: {$day['conditions']}, {$day['temperature_low']}°C to {$day['temperature_high']}°C\n";
}

Schema Validation

With JSON Schema mode, Polyglot ensures the LLM's response adheres to your schema:

1. The schema is sent to the model as part of the request
2. The model structures its response to match the schema
3. For providers with native schema support, validation happens at the API level
4. For other providers, Polyglot helps guide the model to produce correctly formatted output

Provider Support for JSON Schema

Provider support for JSON Schema varies:

• OpenAI (GPT-4 and newer): Native support with json_schema response format
• Anthropic (Claude 3 and newer): Partial support via prompt engineering
• Other providers: May require more explicit instructions in the prompt

For best compatibility, use OpenAI for schema-validated responses.

When to Use JSON Schema Mode

JSON Schema mode is ideal for: - Applications requiring strictly typed data - Integration with databases or APIs that expect specific structures - Data extraction with complex nested structures - Ensuring consistent response formats across multiple requests