-
What is TOON? TOON (Token-Oriented Object Notation) is a lightweight, schema-aware data serialization format specifically designed for Large Language Models (LLMs). Often referred to as "JSON for AI," TOON rethinks how data should be structured to minimize token usage while maintaining maximum readability for models like GPT-4, Claude, and Llama.
-
The Problem with JSON in AI While JSON is the industry standard for web development, it is inefficient for AI interactions:
Verbosity: Constant repetition of keys and heavy use of structural characters (, [], ") consume unnecessary tokens.
High Costs: Since API pricing is based on token counts, JSON's overhead directly increases operational costs.
Context Waste: Every redundant character takes up valuable space in the model’s limited context window.
- Key Features of TOON Up to 60% Token Savings: By eliminating redundant punctuation and utilizing a table-like structure for arrays, TOON drastically reduces the footprint of your data.
Model-Centric Design: TOON combines the hierarchical clarity of YAML with the density of CSV, making it easier for LLMs to parse and reason over complex datasets.
Zero-Loss Conversion: It is designed to be fully interoperable with JSON. You can keep your internal logic in JSON and only convert to TOON when communicating with an LLM.
Human-Readable: Unlike binary formats (like Protobuf), TOON remains easy for developers to read and debug.
- Primary Use Cases Prompt Engineering: Injecting large datasets or reference materials into prompts without hitting token limits.
AI Agents: Reducing overhead during tool-calling and multi-step reasoning loops.
RAG (Retrieval-Augmented Generation): Increasing the density of information retrieved from vector databases to fit more context into a single query.
