Markdown JSON mode is a special mode that requests the model to format its response as JSON within a Markdown code block. This is particularly useful for models or providers that don’t have native JSON output support.

​
Using Markdown JSON Mode

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

$inference = new Inference();

// This works with virtually any provider
$response = $inference->with(
    messages: 'List three programming languages and their key features.',
    mode: OutputMode::MdJson
)->asJsonData();

// The model will return JSON wrapped in Markdown, which Polyglot processes for you
foreach ($response['languages'] as $language) {
    echo "{$language['name']} - {$language['paradigm']}\n";
    echo "Key features: " . implode(', ', $language['key_features']) . "\n\n";
}

​
How MdJson Mode Works

  1. Polyglot instructs the model to respond with a JSON object wrapped in a Markdown code block
  2. The model formats its response accordingly (json {...} )
  3. Polyglot extracts the JSON content from the Markdown code block
  4. The JSON is parsed and returned to your application

​
Providing Guidance for MdJson

While MdJson is more flexible across providers, you still need to provide clear instructions: 
<?php
use Cognesy\Polyglot\Inference\Inference;
use Cognesy\Polyglot\Inference\Enums\OutputMode;

$inference = new Inference();

// Include expected format in the prompt
$prompt = <<<EOT
List three programming languages with their key features.
Respond with a JSON object following this structure:
{
  "languages": [
    {
      "name": "Language name",
      "paradigm": "Programming paradigm",
      "year_created": year as number,
      "key_features": ["feature1", "feature2", "feature3"]
    },
  ]
}
EOT;

$response = $inference->with(
messages: $prompt,
mode: OutputMode::MdJson
)->toJson();

// Process as normal JSON

​
When to Use MdJson Mode

MdJson mode is ideal for:
  • Working with providers that don’t have native JSON output
  • Ensuring portability across different providers
  • Getting structured responses from older model versions
  • Fallback option when JSON Schema mode isn’t available