Skip to main content
Provider abstraction has four parts:
  1. Provider builders (LLMProvider, EmbeddingsProvider)
  2. Runtime assembly (InferenceRuntime, EmbeddingsRuntime)
  3. Driver factories (InferenceDriverFactory, EmbeddingsDriverFactory)
  4. Driver + adapter contracts

​
Provider Builders

​
LLMProvider

<?php
use Cognesy\Polyglot\Inference\LLMProvider;

$provider = LLMProvider::using('openai')
    ->withConfigOverrides(['model' => 'gpt-4o']);

$config = $provider->resolveConfig();
// @doctest id="c462"
Common methods:
  • using(...), dsn(...), new(...)
  • withLLMPreset(...)
  • withLLMConfig(...)
  • withConfigOverrides(...)
  • withDsn(...)
  • withModel(...)
  • withDriver(...) (explicit inference driver instance)

​
EmbeddingsProvider

<?php
use Cognesy\Polyglot\Embeddings\EmbeddingsProvider;

$provider = EmbeddingsProvider::using('openai')
    ->withConfigProvider($configProvider);

$config = $provider->resolveConfig();
// @doctest id="2a6a"
Common methods:
  • using(...), dsn(...), new(...)
  • withPreset(...)
  • withConfig(...)
  • withDsn(...)
  • withDriver(...) (explicit embeddings driver instance)

​
Runtime Assembly

<?php
use Cognesy\Polyglot\Inference\InferenceRuntime;

$runtime = InferenceRuntime::fromProvider($provider);
// @doctest id="f784"
Inference runtime factories:
  • fromProvider(...)
  • fromResolver(...)
  • fromConfig(...)
  • using(...)
  • fromDsn(...)
Embeddings runtime provides the same entry points via EmbeddingsRuntime.

​
Driver Contracts

Inference: 
interface CanProcessInferenceRequest {
    public function makeResponseFor(InferenceRequest $request): InferenceResponse;
    public function makeStreamResponsesFor(InferenceRequest $request): iterable;
    public function capabilities(?string $model = null): DriverCapabilities;
}
// @doctest id="1b80"
Embeddings: 
interface CanHandleVectorization {
    public function handle(EmbeddingsRequest $request): HttpResponse;
    public function fromData(array $data): ?EmbeddingsResponse;
}
// @doctest id="90dd"

​
Adapter Contracts

Inference adapter boundary:
  • CanTranslateInferenceRequest::toHttpRequest(...)
  • CanTranslateInferenceResponse::fromResponse(...)
  • CanTranslateInferenceResponse::fromStreamResponses(...)
  • CanTranslateInferenceResponse::toEventBody(...)
Embeddings adapter boundary:
  • EmbedRequestAdapter::toHttpClientRequest(...)
  • EmbedResponseAdapter::fromResponse(...)

​
Base Drivers

BaseInferenceRequestDriver and BaseEmbedDriver implement the transport pipeline:
  1. map domain request to HttpRequest
  2. execute through shared HttpClient
  3. map raw response payload back to domain response
  4. emit events on request/response/failure

​
Driver Factories and Registry

Inference drivers are selected by LLMConfig::$driver in InferenceDriverFactory. Custom drivers can be registered at runtime: 
<?php
use Cognesy\Polyglot\Inference\Inference;

Inference::registerDriver('my-driver', MyInferenceDriver::class);
Inference::unregisterDriver('my-driver');
Inference::resetDrivers();
// @doctest id="5a40"
Embeddings drivers are selected by EmbeddingsConfig::$driver in EmbeddingsDriverFactory. Custom registration is available via: 
<?php
use Cognesy\Polyglot\Embeddings\Embeddings;

Embeddings::registerDriver('my-embed-driver', MyEmbeddingsDriver::class);
// @doctest id="d9a5"