Loading...
Reduced from 165 to 80 tokens
The architecture of TOON is engineered specifically for the attention mechanisms of Large Language Models.
By eliminating redundant JSON syntax like quotes and colons, we reduce the "noise" tokens that models waste attention on, lowering your baseline costs.
Our algorithm detects uniform arrays and compresses them into a headered tabular block, perfect for large datasets and RAG pipelines.
Unlike lossy compression, TOON is a 1-to-1 mapping. You can convert from JSON to TOON for inference and back to JSON with zero data loss.
TOON (Token-Oriented Object Notation) is a data format optimized for LLM token efficiency. It provides the structure of JSON with the density of CSV.
Yes! Models like GPT-4o, Claude 3.5, and Gemini 1.5 Pro are naturally proficient at parsing indented structure, often outperforming JSON in reasoning.
CSV lacks nesting capability. TOON allows you to mix tabular data with rich object nesting seamlessly, which is essential for complex tool-calling.
Users Typically see 40-65% token reduction depending on your object nesting depth, directly slashing your API bills.
Was this tool helpful?