TOON: Token-Oriented Object Notation

Status: 🔄 Development
Goal: LLM-friendly data format
Languages: Python, JavaScript, Go, Rust

Co to jest TOON?

TOON (Token-Oriented Object Notation) to kompaktowy format modelowania danych zaprojektowany specjalnie dla:

1. Efektywnej komunikacji z LLM - minimalizacja tokenów
2. API definitions - zwięzły opis endpointów
3. Schema definitions - kompaktowe typy danych
4. Config files - czytelne dla ludzi i maszyn

Motywacja

JSON i YAML są verbose - przy komunikacji z LLM każdy token kosztuje. TOON redukuje zużycie tokenów o 40-60% przy zachowaniu czytelności.

Porównanie

JSON (47 tokenów):

{
  "user": {
    "id": 123,
    "name": "John Doe",
    "email": "john@example.com",
    "roles": ["admin", "user"]
  }
}

TOON (19 tokenów):

user{id:123 name:"John Doe" email:"john@example.com" roles:[admin user]}

Redukcja: 60%

Składnia

Podstawy

# Komentarz
name:"value"           # String
count:42               # Number
active:true            # Boolean
items:[a b c]          # Array
nested{key:val}        # Object

API Definition

api{
  name:"User API"
  version:"1.0"
  base:"/api/v1"
  
  endpoints[
    {GET /users -> User[] "List all users"}
    {GET /users/:id -> User "Get user by ID"}
    {POST /users <- UserCreate -> User "Create user"}
    {PUT /users/:id <- UserUpdate -> User "Update user"}
    {DELETE /users/:id -> void "Delete user"}
  ]
  
  types{
    User{id:int name:str email:str roles:str[]}
    UserCreate{name:str email:str}
    UserUpdate{name:str? email:str?}
  }
}

Schema Definition

schema User{
  id:int @primary
  name:str @required @min(1) @max(100)
  email:str @email @unique
  roles:str[] @default([user])
  created:datetime @auto
}

SDK Support

Python

from toon import parse, dump

# Parse TOON
data = parse('user{name:"John" age:30}')
print(data)  # {'user': {'name': 'John', 'age': 30}}

# Generate TOON
toon_str = dump({'user': {'name': 'John', 'age': 30}})
print(toon_str)  # user{name:"John" age:30}

JavaScript

import { parse, dump } from 'toon-js';

const data = parse('user{name:"John" age:30}');
console.log(data);  // { user: { name: 'John', age: 30 } }

const toonStr = dump({ user: { name: 'John', age: 30 } });
console.log(toonStr);  // user{name:"John" age:30}

Status Komponentów

Use Cases

1. LLM Prompts

prompt{
  system:"You are a helpful assistant"
  context{user_prefs{lang:pl theme:dark}}
  history[
    {role:user msg:"Hello"}
    {role:assistant msg:"Hi! How can I help?"}
  ]
}

2. API Documentation

endpoint{
  method:POST
  path:"/orders"
  auth:bearer
  body{product_id:int qty:int}
  response{order_id:int status:str}
  errors[400:"Invalid input" 401:"Unauthorized"]
}

3. Configuration

config{
  app{name:"MyApp" env:production}
  db{host:"localhost" port:5432 pool:10}
  cache{driver:redis ttl:3600}
  features{dark_mode:true beta:false}
}

Benchmarks

Token Count Comparison

Parse Performance (ops/sec)

TOON is ~15-20% slower than JSON parsing, but the token savings offset this for LLM use cases.

Roadmap

Q1 2026

• Go SDK release
• OpenAPI import complete
• CLI tool for conversion

Q2 2026

• Rust SDK
• GraphQL schema support
• Protobuf interop

Q3 2026

• Spec v1.1 with extensions
• Tree-sitter grammar
• LSP server

Linki

• Spec: github.com/wronai/toon-spec
• Python: pypi.org/project/toon-parser
• JavaScript: npmjs.com/package/toon-js
• Playground: toon.dev (planned)

---

Ostatnia aktualizacja: 10 lutego 2026