README

A lightweight, BC-stable PHP client for OpenAI-compatible inference servers. Built for local-first AI in production. Zero framework lock-in.

Quickstart

composer require ndrstmr/steg symfony/http-client

use Steg\Factory\StegClientFactory;

$steg = StegClientFactory::fromDsn('vllm://localhost:8000/v1?model=llama-3.3-70b-awq');
echo $steg->ask('Erkläre Photosynthese in Leichter Sprache.');

Why Steg?

Steg is purpose-built for local inference deployments and provides a BC-promise that symfony/ai-platform does not (yet) offer. Ideal as a stable fallback layer in production systems.

Supported Backends

All backends share the same OpenAiCompatibleClient — DSN prefixes are convenience aliases that resolve to the correct base_url and default port.

Usage

Client creation

use Steg\Factory\StegClientFactory;

// DSN (recommended)
$steg = StegClientFactory::fromDsn('vllm://localhost:8000/v1?model=llama-3.3-70b-awq');
$steg = StegClientFactory::fromDsn('ollama://localhost:11434?model=llama3.2');
$steg = StegClientFactory::fromDsn('mock://default?response=Hello+World');

// Array config (e.g. from Symfony parameters)
$steg = StegClientFactory::fromConfig([
    'base_url' => 'http://localhost:8000/v1',
    'model'    => 'llama-3.3-70b-awq',
    'api_key'  => 'EMPTY',
    'timeout'  => 120,
]);

Completion methods

// One-shot: single user prompt
$answer = $steg->ask('What is Leichte Sprache?');

// System + user: most common chat pattern
$answer = $steg->chat(
    system: 'You translate German administrative texts into Leichte Sprache.',
    user: 'Die Bundesregierung hat neue Gesetze beschlossen.',
);

// Full message history
use Steg\Model\ChatMessage;

$answer = $steg->complete([
    ChatMessage::system('You are a helpful assistant.'),
    ChatMessage::user('What is the capital of France?'),
    ChatMessage::assistant('The capital of France is Paris.'),
    ChatMessage::user('And Germany?'),
])->content;

// Streaming
foreach ($steg->stream([ChatMessage::user('Write a poem.')]) as $chunk) {
    echo $chunk->delta;
}

CompletionOptions presets

use Steg\Model\CompletionOptions;

$steg->ask('Generate JSON.', CompletionOptions::precise());        // temperature 0.1
$steg->ask('Write a story.', CompletionOptions::creative());       // temperature 0.9
$steg->ask('Translate.', CompletionOptions::leichteSprache());     // temperature 0.3
$steg->ask('Anything.', CompletionOptions::default());             // temperature 0.7

// Custom (immutable — returns new instance)
$opts = CompletionOptions::default()->withTemperature(0.5)->withMaxTokens(2048);

Server health and model list

if ($steg->isHealthy()) {
    foreach ($steg->listModels() as $model) {
        echo $model->id.PHP_EOL;
    }
}

Exception Handling

use Steg\Exception\ConnectionException;
use Steg\Exception\InferenceException;
use Steg\Exception\ModelNotFoundException;
use Steg\Exception\InvalidResponseException;

try {
    $response = $steg->ask('Hello');
} catch (ConnectionException $e) {
    // Server unreachable or timeout
} catch (ModelNotFoundException $e) {
    // Model not loaded on the server
    echo 'Missing model: '.$e->getModelId();
} catch (InferenceException $e) {
    // Server returned 4xx/5xx
    echo 'HTTP '.$e->getHttpStatusCode();
} catch (InvalidResponseException $e) {
    // Response parsing failed
}

Testing with MockClient

use Steg\Client\MockClient;
use Steg\StegClient;

// Fixed responses, cycling
$client = new StegClient(MockClient::withResponses([
    'First response',
    'Second response',
]));

$client->ask('anything'); // → 'First response'
$client->ask('anything'); // → 'Second response'
$client->ask('anything'); // → 'First response' (loops)

// Dynamic responses via callback
$client = new StegClient(MockClient::withCallback(
    static fn (array $messages) => 'Echo: '.$messages[0]->content,
));

Symfony Integration

Install the optional bundle for automatic DI configuration and a Symfony Profiler panel:

composer require ndrstmr/steg-bundle

# config/packages/steg.yaml
steg:
    connections:
        vllm_local:
            dsn: '%env(STEG_VLLM_DSN)%'
            timeout: 120
    default_connection: vllm_local

use Steg\Client\InferenceClientInterface;

final class MyService
{
    public function __construct(
        private readonly InferenceClientInterface $steg,
    ) {}
}

See docs/symfony-integration.md for full details.

Requirements

• PHP 8.4+
• psr/log: ^3.0
• symfony/http-client-contracts: ^3.0
• symfony/http-client: 7.4.* (runtime, recommended)

Documentation

• Getting Started
• Configuration Reference
• Supported Backends
• Symfony Integration

License

Licensed under the European Union Public Licence v1.2 (EUPL-1.2).

Origin

Steg was built in a public sector context to solve a real problem: a stable, local LLM client for production use — without framework lock-in.

Built by 👾 public sector dev crew

Notice

This repository was developed with the assistance of AI code agents (Claude Code, Anthropic). The code was created as part of a development sprint and is not cleared for production use without prior review. Use at your own risk.

License: European Union Public Licence v. 1.2 (EUPL-1.2) — Copyright © 2026 Andreas Teumer