README

smartbrew.ai platform bridge for Symfony AI.

Smartbrew exposes an OpenAI-compatible chat API in front of an Ollama backend, so this bridge combines an OpenAI-shaped request/response contract with Ollama-style option handling and NDJSON streaming.

This bridge includes support for:

Chat completions (POST api/chat/completions)
Embeddings (POST api/embeddings)
Streaming responses (NDJSON deltas)
Reasoning / thinking output
Tool calling
Structured output (JSON schema)
Dynamic model discovery (GET api/models)

Installation

composer require sudhaus7/ai-smartbrew-platform

Requires PHP 8.2+, symfony/ai-platform ^0.12 and symfony/http-client ^7.3|^8.0.

Set your API key in the environment so the bridge can pick it up without extra wiring:

SMARTBREW_API_KEY=…

Usage

use Symfony\AI\Platform\Bridge\Smartbrew\Factory;
use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;

$platform = Factory::createPlatform();

$result = $platform->invoke('openai/gpt-oss', new MessageBag(
    Message::forSystem('You are a helpful assistant.'),
    Message::ofUser('What is Symfony AI?'),
));

echo $result->asText();

The endpoint is fixed at https://chat.smartbrew.ai/ — the factory takes no $endpoint argument.

Factory::createProvider() returns the bare provider if you want to compose it with other providers in your own Platform instance. Both factory methods accept an $apiKey, a custom HttpClientInterface, Contract and EventDispatcherInterface.

Authentication

The API key is sent as Authorization: Bearer …. If you do not pass apiKey explicitly, the bridge falls back to $_ENV['SMARTBREW_API_KEY']:

$platform = Factory::createPlatform();                       // from SMARTBREW_API_KEY
$platform = Factory::createPlatform(apiKey: 'sk-…');         // explicit, wins over the env var

Factory::createHttpClient() exposes the same wiring on its own — a base-URI-scoped EventSourceHttpClient with the bearer token applied — if you need a preconfigured client outside the platform, e.g. to build a standalone ModelCatalog:

$catalog = new ModelCatalog();                                    // client built from SMARTBREW_API_KEY
$catalog = new ModelCatalog(Factory::createHttpClient(apiKey: 'sk-…'));

Without a client and without SMARTBREW_API_KEY in the environment, ModelCatalog throws an InvalidArgumentException.

Streaming

Streaming is opt-in — the bridge sends stream: false unless you ask for it:

$result = $platform->invoke('openai/gpt-oss', $messages, ['stream' => true]);

foreach ($result->asStream() as $delta) {
    echo $delta;
}

Text, thinking and completed tool calls are emitted as separate deltas, followed by token usage and a finish_reason metadata delta.

Options

Option handling follows Ollama's split between top-level request keys and nested model options. Known top-level keys (stream, format, keep_alive, tools, think, logprobs, top_logprobs for chat; truncate, keep_alive, dimensions for embeddings) stay at the root of the payload, everything else is moved into options:

$platform->invoke('openai/gpt-oss', $messages, [
    'think' => true,       // top level
    'temperature' => 0.2,  // → options.temperature
]);

Models

The model catalog is resolved at runtime from GET api/models — there is no hardcoded model list. Capabilities are derived from the server's model metadata:

Smartbrew / Ollama capability	Symfony AI capability
`completion`	`INPUT_MESSAGES`
`embedding`	`EMBEDDINGS`
`tools`, `builtin_tools`	`TOOL_CALLING`
`thinking`	`THINKING`
`vision`	`INPUT_IMAGE`

Models that report completion additionally get OUTPUT_TEXT and OUTPUT_STREAMING, and every non-embedding model gets OUTPUT_STRUCTURED. Presets (models flagged preset: true) are read from info.meta.capabilities, plain models from ollama.capabilities. A model whose metadata carries no capabilities raises an InvalidArgumentException — that usually means the Smartbrew server is too old and needs an upgrade.

Resources

License

MIT

sudhaus7 / ai-smartbrew-platform

Maintainers

Package info

Statistics

Security