Skip to main content

​
Overview

Some use cases require tracking the token usage of the API responses. This can be done by getting InferenceUsage object from Instructor LLM response object. Code below demonstrates how it can be retrieved for both sync and streamed requests.

​
Example

<?php
require 'examples/boot.php';

use Cognesy\Instructor\StructuredOutput;
use Cognesy\Polyglot\Inference\Data\InferenceUsage;

class User {
    public int $age;
    public string $name;
}

function printUsage(InferenceUsage $usage) : void {
    echo "Input tokens: $usage->inputTokens\n";
    echo "Output tokens: $usage->outputTokens\n";
    echo "Cache creation tokens: $usage->cacheWriteTokens\n";
    echo "Cache read tokens: $usage->cacheReadTokens\n";
    echo "Reasoning tokens: $usage->reasoningTokens\n";
}

echo "COUNTING TOKENS FOR SYNC RESPONSE\n";
$text = "Jason is 25 years old and works as an engineer.";
$response = StructuredOutput::using('openai')
    ->with(
        messages: $text,
        responseModel: User::class,
    )->response();

echo "\nTEXT: $text\n";
assert($response->usage()->total() > 0);
printUsage($response->usage());


echo "\n\nCOUNTING TOKENS FOR STREAMED RESPONSE\n";
$text = "Anna is 19 years old.";
$stream = StructuredOutput::using('openai')
    ->with(
        messages: $text,
        responseModel: User::class,
        options: ['stream' => true],
    )
    ->stream();

$response = $stream->finalValue();
echo "\nTEXT: $text\n";
printUsage($stream->usage());
?>