mini\Inference\InferenceServiceInterface abstract interface

Documentation

Inference service interface for LLM-based structured evaluation

Applications implement this interface to provide LLM inference capabilities. The evaluate() method takes a prompt and a JSON Schema (via Validator or array) and returns a response guaranteed to match the schema.

Example implementation using Ollama:

class OllamaInference implements InferenceServiceInterface
{
    public function __construct(
        private string $model = 'llama3.2',
        private string $baseUrl = 'http://localhost:11434'
    ) }

    public function evaluate(string $prompt, Validator|\JsonSerializable|array $schema): mixed
    {
        $response = $this->request('/api/generate', [
            'model' => $this->model,
            'prompt' => $prompt,
            'format' => $schema,
            'stream' => false,
        ]);
        return json_decode($response['response'], true);
    }
}

Usage:

use function mini\inference;
use function mini\validator;

// Boolean evaluation
$needsReview = inference()->evaluate(
    "Does this message require human action?\n\n$text",
    validator()->enum([true, false])
);

// Classification
$sentiment = inference()->evaluate(
    "Classify the sentiment:\n\n$text",
    validator()->enum(['positive', 'negative', 'neutral'])
);

// Structured extraction
$data = inference()->evaluate(
    "Extract contact info:\n\n$text",
    validator()->type('object')->properties([
        'name' => validator()->type('string')->required(),
        'email' => validator()->type('string'),
    ])
);

Methods (2)

Source

src/Inference/InferenceServiceInterface.php:61-96