Awesome

⏳ Tiktoken for PHP

Most compatible PHP implementation of OpenAI's original Tiktoken.

Get Started

Requires PHP 8.1+

Install Tiktoken via Composer package manager.

composer require rahul900day/tiktoken-php

Supported Models

Usage

Basic Usage

use Rahul900day\Tiktoken\Tiktoken;

$encoder = Tiktoken::getEncodingForModel('gpt-4');
$encoder->encode("hello world aaaaaaaaaaaa");
$encoder->decode([9906, 4435]);

Special Tokens

use Rahul900day\Tiktoken\Tiktoken;

$encoder = Tiktoken::getEncodingForModel('gpt-4');
$encoder->encode('<|endoftext|>', allowedSpecial: 'all');

Caching

Tiktoken always cache the server's responses when downloading them.

By default it uses the system's default temporary directory to cache a response but you can still overwrite the cache location by setting TIKTOKEN_CACHE_DIR environment variable.

Registering Custom Encoding

use Rahul900day\Tiktoken\Encodings\OpenAiPublic\Cl100KBaseEncoding;

class Cl100KIm extends Cl100KBaseEncoding 
{
    protected function getName(): string
    {
        return 'cl100k_im';
    }
    
    protected function getSpecialTokens(): array
    {
        return [
            ...parent::getSpecialTokens(),
            "<|im_start|>" => 100264,
            "<|im_end|>" => 100265,
        ];
    }
}

use Rahul900day\Tiktoken\Registry;
use Rahul900day\Tiktoken\Tiktoken;

Registry::registerCustomEncoding('cl100k_im', new Cl100KIm);
$encoding = Tiktoken::getEncoding('cl100k_im');

// Expect: 100264
$encoding->encode("<|im_start|>", allowedSpecial: 'all');

Changelog

Please see CHANGELOG for more information on what has changed recently.

Security Vulnerabilities

Please review our security policy on how to report security vulnerabilities.

Credits

• Rahul Dey
• All Contributors

License

This package is released under the MIT License.