When working with LLMs, most developers send structured data as JSON.
It's standard, familiar, and perfect for APIs — but not always efficient for language models.
LLMs don't parse JSON like software does.
They simply read it as text, and we pay for every token — quotes, brackets, spaces, commas, everything.
Result?
Unnecessary cost + slower responses.
This is exactly why TOON (Token-Oriented Object Notation) was created.
What is TOON?
TOON = a compact data format designed for LLM input.
It aims to:
- reduce token usage
- remain human-readable
- preserve structure without heavy syntax
TOON borrows ideas from:
| Format | Borrowed Concept |
|---|---|
| YAML | indentation for structure |
| CSV | table-like lists for repeated fields |
The goal: minimum characters, maximum clarity.
Example
JSON
{ "tags": ["chill", "lofi", "study"] }
TOON
tags(3) - chill - lofi - study
Same meaning.
Fewer tokens.
Why TOON Matters
| Benefit | Impact |
|---|---|
| Token-efficient | ~30–60% fewer tokens in tests |
| Faster prompts | Less text to process |
| Readable | Easier debugging + tracing |
| LLM-aware format | Designed for prompt input, not APIs |
Token Comparison (Real Test)
Flat user dataset → query LLM
| Format | Prompt Tokens | Total Tokens |
|---|---|---|
| JSON | 757 | 797 |
| TOON | 460 | 500 ✅ |
~40% reduction with zero data loss.
Example Workflow
JSON approach
const input = JSON.stringify(data)
sendToLLM(input)
TOON approach
import { encode } from "@microsoft/tone"
const input = encode(data)
sendToLLM(input)
Clearer. Smaller. Faster.
⚠️ Important — TOON Works Best With Flat Data
- Flat JSON → TOON ✅ efficient
- Nested JSON → TOON ❌ can cost more tokens
If your data has nested objects, flatten first before encoding.
When to Use TOON
Use TOON for:
✅ passing structured data in prompts
✅ RAG agents and memory context
✅ large dynamic input payloads
✅ debugging structured prompt data
Don’t use TOON for:
❌ APIs
❌ storage
❌ LLM output format
Final Thoughts
TOON isn’t replacing JSON — both have roles.
| JSON | TOON |
|---|---|
| Best for machines & APIs | Best for LLM input payloads |
| Heavy syntax | Lightweight |
| Universal | Optimized for tokens |
As LLM workloads scale, token efficiency becomes engineering.
TOON is a practical optimization worth exploring for real-world AI development.
Top comments (0)