ASON — Array-Schema Object Notation

ASON is a compact, schema-driven data format designed for LLM interactions and high-performance data transmission. It separates schema from data — keys are declared once, rows carry only values.

中文文档

---

Why ASON?

Standard JSON repeats every field name for every record. When sending structured datasets to an LLM or over a network, that redundancy burns tokens and bandwidth:

[
  { "id": 1, "name": "Alice", "active": true },
  { "id": 2, "name": "Bob", "active": false },
  { "id": 3, "name": "Carol", "active": true }
]

ASON declares the schema once and streams data as compact tuples:

[{id:int, name:str, active:bool}]:(1,Alice,true),(2,Bob,false),(3,Carol,true)

~65% fewer tokens. Same information. Machine-readable schema.

---

ASON vs JSON

Aspect	JSON	ASON
Token efficiency	100% (baseline)	30–70% ✓
Key repetition	Every object	Declared once ✓
Type annotations	None	Optional ✓
Human readable	Yes	Yes ✓
Nested structs	✓	✓
Serialization speed	1×	~1.7–2.4× faster ✓
Deserialization speed	1×	~1.9–2.9× faster ✓
Data size	100% (baseline)	40–60% ✓
Binary codec	✗	✓
Struct reflection	✗	Compile-time ✓

Token Savings — A Concrete Example

JSON (100 tokens):
{"users":[{"id":1,"name":"Alice","active":true},{"id":2,"name":"Bob","active":false}]}

ASON (~35 tokens, 65% saving):
[{id:int, name:str, active:bool}]:(1,Alice,true),(2,Bob,false)

The schema header also serves as an inline hint for LLMs — field names and optional types are immediately visible without scanning every row.

---

ASON vs TOON

TOON (Token-Oriented Object Notation) is another format designed for reducing tokens in LLM prompts. Both ASON and TOON share the core idea of eliminating key repetition for array-of-object data, but they differ significantly in design goals and scope.

Syntax at a glance

TOON — indentation-based, YAML-inspired:

users[2]{id,name,active}:
  1,Alice,true
  2,Bob,false

ASON — tuple-based, schema-explicit:

[{id:int, name:str, active:bool}]:(1,Alice,true),(2,Bob,false)

Comparison Table

Aspect	TOON	ASON
Schema declaration	Auto-detected at encode time	Explicit, reusable, language-level ✓
Type annotations	None (JSON data model only)	Optional rich types (int, str, bool, f64, opt_*, vec_*, map_*) ✓
Syntax style	YAML-like indentation	Compact tuple rows
Array length markers	[N] — helps detect truncation	Schema header defines structure ✓
Nested structures	Falls back to verbose list format	Native, efficient, recursive ✓
Use case	LLM input only (read-only layer)	LLM + serialization + data transmission ✓
Serialization speed	Not measured (JS only)	SIMD-accelerated, ~1.7–2.9× vs JSON ✓
Deserialization speed	Not measured (JS only)	Zero-copy parsing ✓
Binary codec	✗	✓
Language implementations	TypeScript / JavaScript only	C, C++, C#, Go, Java, JS, Python, Rust, Zig, Dart ✓
Struct reflection	Dynamic (runtime only)	Compile-time (ASON_FIELDS macro) ✓
Deeply nested data	Token cost increases significantly	Efficient at any nesting level ✓
Round-trip fidelity	JSON data model (no type info)	Full type fidelity ✓

When to Choose ASON

• You need high-performance serialization in a compiled language (C, C++, Rust, Go, …)
• Your data has rich types — optional fields, typed vectors, maps, nested structs
• You need binary encoding alongside the text format
• You work in multiple languages or need a language-neutral wire format
• You want the schema to act as a self-documenting API contract for LLM prompts

When TOON May Be Enough

• Your pipeline is LLM prompt input only (you don't parse it back into structs)
• Your data is simple flat tables with no type constraints

---

Format Overview

Single Object

{id:int, name:str, active:bool}:(42,Alice,true)

Array of Objects (Schema-Driven)

Schema declared once, each row is a tuple:

[{id:int, name:str, active:bool}]:(1,Alice,true),(2,Bob,false),(3,Carol,true)

Nested Structs

{name:str, dept}:(Alice,(Engineering))

Optional Fields

{id:int, label:opt_str}:(1,hello),(2,)

(blank value = None/null)

Typed Vectors and Maps

{name:str, tags:vec_str, attrs:map_si}:(Alice,[rust,go,python],[(age,30),(score,95)])

---

Implementations

Language	Repository	Notes
C	ason-c	✓ C11
C++	ason-cpp	✓ C++17
C#	ason-cs	✓ .NET 8+
Go	ason-go	✓ Go 1.24+
Java / Kotlin	ason-java	✓ Java 21+
JavaScript / TypeScript	ason-js	✓ JS/TS
Python	ason-py	✓
Rust	ason-rs	✓
Zig	ason-zig	✓
Dart	ason-dart	✓
PHP	ason-php	✓ PHP 8.4+
Swift	ason-swift	TODO

Plugins

IDE	Repository	Notes
VSCode	plugin_vscode	✓
Jetbrain	plugin_jetbrain	Todo
Zed	plugin_zed	Todo

---

License

MIT