toon is a dependency-free R package that provides functions to serialize native R objects (lists, data frames, vectors) into the Token-Oriented Object Notation (TOON) format.

TOON is specifically designed to significantly reduce token usage when passing structured data to Large Language Models (LLMs). It offers a substantial token-efficiency advantage over traditional formats like JSON, particularly for tabular or highly uniform data.

Install the stable or development versions:

# CRAN version
install.packages("toon")
# Dev version
## If you don't have 'remotes' installed, run: install.packages('remotes')
remotes::install_github("laresbernardo/toon")

The primary function is as_toon(), which takes an R object and returns a toon class string.

Let's define a complex R list containing vectors, a data frame, and nested lists (objects and arrays):

Config_Test <- list(
  Top_Array = c("alpha", "beta", "gamma"),
  Empty_Settings = list(),
  User_Profile = list(
    ID = 12345,
    Is_Admin = TRUE,
    Status = "Active",
    Null_Field = NULL,
    Not_A_Number = NaN
  ),
  Usage_Log = data.frame(
    time = c(9.1, 15.4),
    action = c("login", "update"),
    success = c(TRUE, FALSE),
    stringsAsFactors = FALSE
  ),
  Servers = list(
    list(
      host = "web01.int",
      ip = "10.0.0.1",
      ports = c(80, 443)
    ),
    c("failover.ext", "192.168.1.1"),
    list(
      host = "db01.int",
      locked = TRUE
    )
  )
)

Calling as_toon(Config_Test) yields the following token-optimized format:

Top_Array: [3]: "alpha","beta","gamma"
Empty_Settings: []
User_Profile:
  ID: 12345
  Is_Admin: true
  Status: "Active"
  Null_Field: null
  Not_A_Number: null
Usage_Log:
[2]{time,action,success}:
  9.1,"login",true
  15.4,"update",false
Servers:
[3]:
  - host: "web01.int"
    ip: "10.0.0.1"
    ports: [2]: 80,443
  - [2]: "failover.ext","192.168.1.1"
  - host: "db01.int"
    locked: true

While the complex example shows toon's restructuring capabilities, its most powerful use case is dramatically reducing the size of datasets.

x <- iris # A simple 150 rows and 5 columns dataset
# Compare the TOON string size to a standard JSON string
nchar(as_toon(x)) / nchar(jsonlite::toJSON(x)) 
# Result: ~0.28 (Same info, roughly 28% the size)

The toon package is based on and inspired by the official format specification. You can read more about the Token-Oriented Object Notation here:

toon-format/toon