Skip to main content

​
Overview

This example uses InferenceRuntime directly and sends the LLM and HTTP lifecycle to Langfuse. It is useful when you want visibility into the raw LLM call path without the additional StructuredOutput layer. Key concepts:
  • InferenceRuntime: direct Polyglot runtime for inference calls
  • PolyglotTelemetryProjector: maps inference lifecycle events
  • HttpClientTelemetryProjector: captures transport spans
  • Telemetry::flush(): pushes the final batch to Langfuse

​
Example

<?php
require 'examples/boot.php';
require_once 'examples/_support/langfuse.php';

use Cognesy\Events\Dispatchers\EventDispatcher;
use Cognesy\Http\Telemetry\HttpClientTelemetryProjector;
use Cognesy\Messages\Messages;
use Cognesy\Polyglot\Inference\Inference;
use Cognesy\Polyglot\Inference\InferenceRuntime;
use Cognesy\Polyglot\Inference\LLMProvider;
use Cognesy\Polyglot\Telemetry\PolyglotTelemetryProjector;
use Cognesy\Telemetry\Application\Projector\CompositeTelemetryProjector;
use Cognesy\Telemetry\Application\Projector\RuntimeEventBridge;

$events = new EventDispatcher('examples.b03.telemetry-langfuse');
$hub = exampleLangfuseHub();

(new RuntimeEventBridge(new CompositeTelemetryProjector([
    new PolyglotTelemetryProjector($hub),
    new HttpClientTelemetryProjector($hub),
])))->attachTo($events);

$runtime = InferenceRuntime::fromProvider(
    provider: LLMProvider::using('openai'),
    events: $events,
);

$response = Inference::fromRuntime($runtime)
    ->with(
        messages: Messages::fromString('Summarize why observability matters for LLM applications in exactly 3 bullet points.'),
        options: ['max_tokens' => 180],
    )
    ->response();

$hub->flush();

echo "Response:\n";
echo $response->content() . "\n\n";
if ($response->usage() !== null) {
    echo "Tokens: {$response->usage()->inputTokens} in / {$response->usage()->outputTokens} out\n";
}
echo "Telemetry: flushed to Langfuse\n";

assert($response->content() !== '');
?>