Skip to main content

​
InferenceRequest

InferenceRequest is the canonical payload passed to inference drivers. 
<?php
use Cognesy\Polyglot\Inference\Data\InferenceRequest;
use Cognesy\Polyglot\Inference\Enums\OutputMode;

$request = new InferenceRequest(
    messages: 'Summarize this text',
    model: 'gpt-4o-mini',
    options: ['temperature' => 0.2],
    mode: OutputMode::Text,
);
// @doctest id="9bce"
Most used accessors:
  • messages()
  • model()
  • tools(), toolChoice()
  • responseFormat()
  • options()
  • outputMode()
  • cachedContext()
  • responseCachePolicy()
  • retryPolicy()
Most used mutators:
  • withMessages(...)
  • withModel(...)
  • withStreaming(...)
  • withTools(...)
  • withToolChoice(...)
  • withResponseFormat(...)
  • withOptions(...)
  • withOutputMode(...)
  • withCachedContext(...)
  • withResponseCachePolicy(...)
  • withRetryPolicy(...)

​
PendingInference

PendingInference defers execution until you read output. 
<?php
use Cognesy\Polyglot\Inference\Inference;

$pending = Inference::using('openai')
    ->withMessages('Explain event sourcing in one paragraph.')
    ->create();

$text = $pending->get();
$json = $pending->asJsonData();
$response = $pending->response();
// @doctest id="80e8"
Core methods:
  • get(): text content
  • asJson(), asJsonData()
  • response(): full InferenceResponse
  • stream(): InferenceStream (only when streaming is enabled)
  • isStreamed()

​
InferenceStream

InferenceStream exposes partial snapshots during streaming. 
<?php
$stream = Inference::using('openai')
    ->withMessages('Write three short title ideas.')
    ->withStreaming(true)
    ->stream();

foreach ($stream->responses() as $partial) {
    echo $partial->contentDelta;
}

$final = $stream->final();
// @doctest id="1778"
Core methods:
  • responses(): generator of PartialInferenceResponse
  • all(): collect all partial responses
  • final(): materialized final InferenceResponse
  • onPartialResponse(callable)
  • map(...), filter(...), reduce(...)

​
PartialInferenceResponse

Each streamed chunk is represented as a cumulative snapshot:
  • contentDelta / content()
  • reasoningContentDelta / reasoningContent()
  • toolId, toolName, toolArgs
  • toolCalls()
  • finishReason()
  • usage()

​
EmbeddingsRequest and EmbeddingsResponse

Embeddings follow the same deferred pattern. 
<?php
use Cognesy\Polyglot\Embeddings\Embeddings;

$pending = Embeddings::using('openai')
    ->withInputs(['doc one', 'doc two'])
    ->create();

$response = $pending->get();
$first = $response->first();
$vectors = $response->vectors();
$usage = $response->usage();
// @doctest id="1ef2"
EmbeddingsResponse also provides last(), split(...), toValuesArray(), and toArray().

​
Identity Types

IDs are value objects serialized as strings at boundaries:
  • InferenceRequestId
  • InferenceExecutionId
  • InferenceAttemptId
  • InferenceResponseId
  • PartialInferenceResponseId
  • ToolCallId