tiktoken

What is tiktoken?

The tiktoken npm package is designed for tokenizing text, particularly for use with OpenAI's GPT models. It provides efficient and accurate tokenization, which is essential for natural language processing tasks.

What are tiktoken's main functionalities?

Tokenization

This feature allows you to tokenize a string of text into tokens. The example demonstrates how to encode a simple string using the GPT-3 encoding.

const tiktoken = require('tiktoken');
const encoder = tiktoken.getEncoding('gpt-3');
const tokens = encoder.encode('Hello, world!');
console.log(tokens);

Detokenization

This feature allows you to convert tokens back into the original text. The example shows how to decode tokens back into the original string.

const tiktoken = require('tiktoken');
const encoder = tiktoken.getEncoding('gpt-3');
const tokens = encoder.encode('Hello, world!');
const text = encoder.decode(tokens);
console.log(text);

Custom Encoding

This feature allows you to create a custom encoding scheme. The example demonstrates how to define a custom encoding and use it to tokenize a string.

const tiktoken = require('tiktoken');
const customEncoding = tiktoken.createEncoding({
  'Hello': 1,
  'world': 2,
  '!': 3
});
const tokens = customEncoding.encode('Hello, world!');
console.log(tokens);

Other packages similar to tiktoken

tokenizer

The tokenizer package provides basic tokenization functionalities. It is more general-purpose compared to tiktoken, which is specifically optimized for OpenAI's GPT models.

natural

The natural package is a comprehensive natural language processing library for Node.js. It includes tokenization as one of its many features, making it more versatile but potentially less optimized for specific use cases like tiktoken.

wink-tokenizer

The wink-tokenizer package is a fast and lightweight tokenizer for JavaScript. It offers similar tokenization capabilities but lacks the specific optimizations for GPT models that tiktoken provides.

README

⏳ tiktoken

tiktoken is a BPE tokeniser for use with OpenAI's models, forked from the original tiktoken library to provide JS/WASM bindings for NodeJS and other JS runtimes.

This repository contains the following packages:

• tiktoken (formally hosted at @dqbd/tiktoken): WASM bindings for the original Python library, providing full 1-to-1 feature parity.
• js-tiktoken: Pure JavaScript port of the original library with the core functionality, suitable for environments where WASM is not well supported or not desired (such as edge runtimes).

Documentation for js-tiktoken can be found in here. Documentation for the tiktoken can be found here below.

The WASM version of tiktoken can be installed from NPM:

npm install tiktoken

Usage

Basic usage follows, which includes all the OpenAI encoders and ranks:

import assert from "node:assert";
import { get_encoding, encoding_for_model } from "tiktoken";

const enc = get_encoding("gpt2");
assert(
  new TextDecoder().decode(enc.decode(enc.encode("hello world"))) ===
    "hello world"
);

// To get the tokeniser corresponding to a specific model in the OpenAI API:
const enc = encoding_for_model("text-davinci-003");

// Extend existing encoding with custom special tokens
const enc = encoding_for_model("gpt2", {
  "<|im_start|>": 100264,
  "<|im_end|>": 100265,
});

// don't forget to free the encoder after it is not used
enc.free();

In constrained environments (eg. Edge Runtime, Cloudflare Workers), where you don't want to load all the encoders at once, you can use the lightweight WASM binary via tiktoken/lite.

const { Tiktoken } = require("tiktoken/lite");
const cl100k_base = require("tiktoken/encoders/cl100k_base.json");

const encoding = new Tiktoken(
  cl100k_base.bpe_ranks,
  cl100k_base.special_tokens,
  cl100k_base.pat_str
);
const tokens = encoding.encode("hello world");
encoding.free();

If you want to fetch the latest ranks, use the load function:

const { Tiktoken } = require("tiktoken/lite");
const { load } = require("tiktoken/load");
const registry = require("tiktoken/registry.json");
const models = require("tiktoken/model_to_encoding.json");

async function main() {
  const model = await load(registry[models["gpt-3.5-turbo"]]);
  const encoder = new Tiktoken(
    model.bpe_ranks,
    model.special_tokens,
    model.pat_str
  );
  const tokens = encoder.encode("hello world");
  encoder.free();
}

main();

If desired, you can create a Tiktoken instance directly with custom ranks, special tokens and regex pattern:

import { Tiktoken } from "../pkg";
import { readFileSync } from "fs";

const encoder = new Tiktoken(
  readFileSync("./ranks/gpt2.tiktoken").toString("utf-8"),
  { "<|endoftext|>": 50256, "<|im_start|>": 100264, "<|im_end|>": 100265 },
  "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)|\\s+"
);

Finally, you can a custom init function to override the WASM initialization logic for non-Node environments. This is useful if you are using a bundler that does not support WASM ESM integration.

import { get_encoding, init } from "tiktoken/init";

async function main() {
  const wasm = "..."; // fetch the WASM binary somehow
  await init((imports) => WebAssembly.instantiate(wasm, imports));

  const encoding = get_encoding("cl100k_base");
  const tokens = encoding.encode("hello world");
  encoding.free();
}

main();

Compatibility

As this is a WASM library, there might be some issues with specific runtimes. If you encounter any issues, please open an issue.

Runtime	Status	Notes
Node.js	✅
Bun	✅
Vite	✅	See here for notes
Next.js	✅	See here for notes
Create React App (via Craco)	✅	See here for notes
Vercel Edge Runtime	✅	See here for notes
Cloudflare Workers	✅	See here for notes
Deno	❌	Currently unsupported (see dqbd/tiktoken#22)
Svelte + Cloudflare Workers	❌	Currently unsupported (see dqbd/tiktoken#37)

For unsupported runtimes, consider using js-tiktoken, which is a pure JS implementation of the tokeniser.

Vite

If you are using Vite, you will need to add both the vite-plugin-wasm and vite-plugin-top-level-await. Add the following to your vite.config.js:

import wasm from "vite-plugin-wasm";
import topLevelAwait from "vite-plugin-top-level-await";
import { defineConfig } from "vite";

export default defineConfig({
  plugins: [wasm(), topLevelAwait()],
});

Next.js

Both API routes and /pages are supported with the following next.config.js configuration.

// next.config.json
const config = {
  webpack(config, { isServer, dev }) {
    config.experiments = {
      asyncWebAssembly: true,
      layers: true,
    };

    return config;
  },
};

Usage in pages:

import { get_encoding } from "tiktoken";
import { useState } from "react";

const encoding = get_encoding("cl100k_base");

export default function Home() {
  const [input, setInput] = useState("hello world");
  const tokens = encoding.encode(input);

  return (
    <div>
      <input
        type="text"
        value={input}
        onChange={(e) => setInput(e.target.value)}
      />
      <div>{tokens.toString()}</div>
    </div>
  );
}

Usage in API routes:

import { get_encoding } from "tiktoken";
import { NextApiRequest, NextApiResponse } from "next";

export default function handler(req: NextApiRequest, res: NextApiResponse) {
  const encoding = get_encoding("cl100k_base");
  const tokens = encoding.encode("hello world");
  encoding.free();
  return res.status(200).json({ tokens });
}

Create React App

By default, the Webpack configugration found in Create React App does not support WASM ESM modules. To add support, please do the following:

1. Swap react-scripts with craco, using the guide found here: https://craco.js.org/docs/getting-started/.
2. Add the following to craco.config.js:

module.exports = {
  webpack: {
    configure: (config) => {
      config.experiments = {
        asyncWebAssembly: true,
        layers: true,
      };

      // turn off static file serving of WASM files
      // we need to let Webpack handle WASM import
      config.module.rules
        .find((i) => "oneOf" in i)
        .oneOf.find((i) => i.type === "asset/resource")
        .exclude.push(/\.wasm$/);

      return config;
    },
  },
};

Vercel Edge Runtime

Vercel Edge Runtime does support WASM modules by adding a ?module suffix. Initialize the encoder with the following snippet:

// @ts-expect-error
import wasm from "tiktoken/lite/tiktoken_bg.wasm?module";
import model from "tiktoken/encoders/cl100k_base.json";
import { init, Tiktoken } from "tiktoken/lite/init";

export const config = { runtime: "edge" };

export default async function (req: Request) {
  await init((imports) => WebAssembly.instantiate(wasm, imports));

  const encoding = new Tiktoken(
    model.bpe_ranks,
    model.special_tokens,
    model.pat_str
  );

  const tokens = encoding.encode("hello world");
  encoding.free();

  return new Response(`${tokens}`);
}

Cloudflare Workers

Similar to Vercel Edge Runtime, Cloudflare Workers must import the WASM binary file manually and use the tiktoken/lite version to fit the 1 MB limit. However, users need to point directly at the WASM binary via a relative path (including ./node_modules/).

Add the following rule to the wrangler.toml to upload WASM during build:

[[rules]]
globs = ["**/*.wasm"]
type = "CompiledWasm"

Initialize the encoder with the following snippet:

import { init, Tiktoken } from "tiktoken/lite/init";
import wasm from "./node_modules/tiktoken/lite/tiktoken_bg.wasm";
import model from "tiktoken/encoders/cl100k_base.json";

export default {
  async fetch() {
    await init((imports) => WebAssembly.instantiate(wasm, imports));
    const encoder = new Tiktoken(
      model.bpe_ranks,
      model.special_tokens,
      model.pat_str
    );
    const tokens = encoder.encode("test");
    encoder.free();
    return new Response(`${tokens}`);
  },
};

Acknowledgements

• https://github.com/zurawiki/tiktoken-rs