Adapters

Polyglot drivers are composed from small, focused adapter classes. Each adapter handles one aspect of the translation between Polyglot’s unified data model and a provider’s native HTTP format. This composition makes it straightforward to add new providers — most of the logic is shared, and only the provider-specific differences need new code.

Adapter Responsibilities

Every inference driver is built from two main translators, each of which may use additional formatters internally:

Request Translation

The request adapter converts a Polyglot InferenceRequest into an HttpRequest. It is responsible for:

• Message formatting — mapping Polyglot’s typed Messages (with roles, content parts, tool calls, and tool results) into the provider’s expected structure
• Body formatting — assembling the full request body including model, tools, response format, and mode-specific adjustments
• HTTP request assembly — setting the URL, headers (including authentication), and body

These responsibilities are typically split across three classes:

Class Pattern	Contract	Purpose
*MessageFormat	CanMapMessages	Maps Messages to provider format
*BodyFormat	CanMapRequestBody	Assembles the full request body
*RequestAdapter	CanTranslateInferenceRequest	Builds the final HttpRequest

Response Translation

The response adapter converts raw HTTP responses back into Polyglot data objects:

Class Pattern	Contract	Purpose
*ResponseAdapter	CanTranslateInferenceResponse	Parses responses and stream deltas
*UsageFormat	CanMapUsage	Extracts token usage from response data

How They Compose

For most providers the wiring is not code at all — it is a row in the bundled registry. An InferenceDriverSpec names the pieces, and SpecifiedInferenceDriver is the single class behind every provider declared this way:

$table = [
    'openai' => new InferenceDriverSpec(bodyFormat: OpenAIBodyFormat::class),
    'groq' => new InferenceDriverSpec(
        bodyFormat: GroqBodyFormat::class,
        usageFormat: GroqUsageFormat::class,
        capabilities: new DriverCapabilities(responseFormatWithTools: false),
    ),
];
// @doctest id="0a27"

Anything left out defaults to the OpenAI implementation, so a provider that differs only in its body format is one line. The spec assembles them in the fixed nesting order every driver used:

$driver = new SpecifiedInferenceDriver(
    config: $config,
    httpClient: $httpClient,
    events: $events,
    requestTranslator: new OpenAIRequestAdapter(
        $config,
        new OpenAIBodyFormat($config, new OpenAIMessageFormat()),
    ),
    responseTranslator: new OpenAIResponseAdapter(new OpenAIUsageFormat()),
    capabilities: null,
);
// @doctest id="aa0b"

Providers that assemble their own URL or headers still use bespoke request adapters, but they are selected by the requestAdapter field in their InferenceDriverSpec; they do not need a provider driver class. The adapter owns that provider-specific behavior while SpecifiedInferenceDriver supplies the shared execution lifecycle.

The BaseInferenceRequestDriver handles the shared execution logic — sending HTTP requests, reading responses, and parsing event streams. The adapters only need to handle format translation.

The Contracts

Request Side

The CanTranslateInferenceRequest contract defines a single method:

interface CanTranslateInferenceRequest
{
    public function toHttpRequest(InferenceRequest $request): HttpRequest;
}
// @doctest id="8c46"

Request adapters typically delegate body construction to a CanMapRequestBody implementation:

interface CanMapRequestBody
{
    public function toRequestBody(InferenceRequest $request): array;
}
// @doctest id="4157"

Message formatting is handled by CanMapMessages, which receives typed Messages and returns a provider-native array. Implementations compose a MessageMapper utility for typed iteration instead of duplicating the loop:

interface CanMapMessages
{
    public function map(Messages $messages): array;
}
// @doctest id="dfcf"

A request adapter composes these together. It receives a CanMapRequestBody (which itself wraps a CanMapMessages) and produces the final HttpRequest.

Almost none of that is per-provider. BaseHttpRequestAdapter owns the skeleton — POST, the body from the body format, the stream flag, the telemetry-correlation wrapper — and leaves two abstract hooks, which are the only things providers actually disagree about:

abstract class BaseHttpRequestAdapter implements CanTranslateInferenceRequest
{
    public function __construct(
        protected LLMConfig $config,
        protected CanMapRequestBody $bodyFormat,
    ) {}

    public function toHttpRequest(InferenceRequest $request): HttpRequest { /* the skeleton */ }

    abstract protected function toUrl(InferenceRequest $request): string;
    abstract protected function toHeaders(InferenceRequest $request): array;
}
// @doctest id="59b7"

So OpenAIRequestAdapter is just the two hooks:

class OpenAIRequestAdapter extends BaseHttpRequestAdapter
{
    protected function toUrl(InferenceRequest $request): string {
        return "{$this->config->apiUrl}{$this->config->endpoint}";
    }

    protected function toHeaders(InferenceRequest $request): array {
        return [
            'Authorization' => "Bearer {$this->config->apiKey}",
            'Content-Type' => 'application/json; charset=utf-8',
            'Accept' => 'application/json',
        ];
    }
}
// @doctest id="8950"

Both hooks are abstract rather than defaulted to {apiUrl}{endpoint}. Only two of the bundled adapters would use such a default; the rest assemble a URL from a region, a model name or a fallback endpoint, and a silently plausible URL is a worse failure than a compile error.

AzureOpenAIRequestAdapter, CohereV2RequestAdapter, GeminiOAIRequestAdapter and HuggingFaceRequestAdapter extend OpenAIRequestAdapter rather than the base directly — they are OpenAI-compatible providers that differ only in headers, and are happy to inherit the URL.

Response Side

The CanTranslateInferenceResponse contract handles both synchronous and streaming responses:

interface CanTranslateInferenceResponse
{
    public function fromResponse(HttpResponse $response): ?InferenceResponse;

    /** @return iterable<PartialInferenceDelta> */
    public function fromStreamDeltas(
        iterable $eventBodies,
        ?HttpResponse $responseData = null,
    ): iterable;

    public function toEventBody(string $data): string|bool;
}
// @doctest id="6a83"

The toEventBody() method extracts the payload from an SSE line (stripping the data: prefix, detecting [DONE] markers). The fromStreamDeltas() method parses a sequence of those payloads into PartialInferenceDelta objects carrying incremental content, tool call fragments, and usage snapshots.

Usage extraction is handled by CanMapUsage:

interface CanMapUsage
{
    public function fromData(array $data): InferenceUsage;
}
// @doctest id="54c6"

Different providers report token usage under different keys and with different granularity. Some include cache tokens or reasoning tokens, others do not. Each provider’s usage formatter encapsulates these differences into the normalized InferenceUsage object.

Streamed usage: pass null, do not build a zero

Usage arrives on only a handful of events per stream — roughly one chunk in 943 on a real one. A response adapter must therefore pass usage: null on every delta that carries no usage payload, rather than calling its usage format unconditionally:

$delta = new PartialInferenceDelta(
    contentDelta: $content,
    usage: $this->hasUsageData($data) ? $this->usageFormat->fromData($data) : null,
);
// @doctest id="dcc4"

StreamingUsageState::apply() already discards a zero-total InferenceUsage, so building one is pure waste — one constructor call and one allocate/free cycle per delta, ten thousand of them on a long stream.

The predicate is provider-specific. Do not copy another adapter’s. CanMapUsage::fromData() implementations read different keys, and a guard that checks the wrong one silently drops real token counts:

Adapter	Usage key(s)
OpenAI, OpenAI-compatible, reasoning variants	usage
Gemini	usageMetadata
Anthropic	usage and message.usage — input tokens arrive on message_start, output tokens on message_delta
Cohere v2	usage and delta.usage
OpenResponses	usage and response.usage — carried on the terminal response.completed event

Express it as a protected function hasUsageData(array $data): bool on the adapter, so the per-provider difference is visible and testable. OpenAIResponseAdapter defines the default (!empty($data['usage'])); subclasses that read usage from elsewhere override it.

Two tests enforce this contract, and a new adapter should be added to both:

• tests/Unit/Drivers/StreamDeltaUsageGuardTest.php — asserts usage === null on a quiet delta, non-null on a carrying one, and that the assembled totals are unchanged.
• tests/Benchmarks/StreamAdapterUsageAllocationTest.php — counts actual InferenceUsage constructions over a 1,000-delta stream and fails if an adapter exceeds O(1).

Varying the response_format payload

Providers disagree sharply about how a response format reaches them: some take a json_schema envelope with a name and a strict flag, some accept only json_object, some hang the schema off json_object or off a bare value key, and some have no JSON-schema support at all and must degrade to plain JSON.

A body format expresses that by overriding one of three methods on OpenAIBodyFormat. Each receives the ResponseFormat the caller asked for and returns the payload to send:

Method	Called for	Base behaviour
toTextResponseFormat()	text	['type' => 'text']
toJsonObjectResponseFormat()	json_object	['type' => 'json_object']
toJsonSchemaResponseFormat()	json_schema	json_schema envelope with name, schema, strict

Override only what differs. A provider with no schema support degrades by delegating:

class MyProviderBodyFormat extends OpenAICompatibleBodyFormat
{
    // Provider accepts json_object only -- schema mode degrades rather than being rejected.
    #[\Override]
    protected function toJsonSchemaResponseFormat(ResponseFormat $responseFormat): array {
        return $this->toJsonObjectResponseFormat($responseFormat);
    }
}
// @doctest id="e58f"

Do not override toResponseFormat() to do this. That method decides whether a response format is sent at all, while the three methods above decide what the payload looks like once that question is settled. DeepSeek V4 uses the latter hook to degrade JSON Schema to JSON Output.

Provider variation used to be injected into ResponseFormat itself through withToTextHandler() / withToJsonObjectHandler() / withToJsonSchemaHandler(). Those methods no longer exist: ResponseFormat is a plain four-field value object and carries no rendering behaviour. Every request used to allocate two closures and two copies of it to reach a payload the body format already had everything to build.

Whatever these return is pinned by ResponseFormatFragmentGoldenTest, which snapshots the emitted fragment for every body format in the family across every mode. A new provider must be added to its table — the test fails if one is missing.

Embeddings Adapters

Embeddings drivers follow the same pattern with their own set of contracts:

Contract	Purpose
EmbedRequestAdapter	Converts EmbeddingsRequest to HttpRequest
EmbedResponseAdapter	Converts HttpResponse to EmbeddingsResponse
CanMapRequestBody	Assembles the embeddings request body
CanMapUsage	Extracts usage from embeddings response data

Adding a New Provider

To add support for a new provider, you typically need to create:

1. A message format class if the provider uses a non-OpenAI message structure
2. A body format class to assemble requests with any provider-specific fields
3. A request adapter to set the URL, headers, and authentication scheme
4. A response adapter to parse responses and streaming events
5. A usage format class if token usage is reported differently
6. An InferenceDriverSpec naming those pieces — or, if the provider assembles its own URL or headers, a driver class extending BaseInferenceRequestDriver

Many providers use OpenAI-compatible formats. In those cases you can reuse the OpenAI adapters directly and skip the driver class entirely: register an InferenceDriverSpec with OpenAICompatibleBodyFormat. The ollama, together, moonshot and openai-compatible names all share exactly that one spec in the bundled registry.