dx-serializer 0.1.0

A token-efficient serialization format for LLM context windows with high-performance binary encoding
Documentation

DX Serializer

Token-optimized serialization format for AI context windows with 52-73% token savings vs JSON and pure RKYV binary format.

Direct JSON vs .machine Proof

DX Serializer's generated .machine files are not a cosmetic cache. When used directly, the typed RKYV + mmap path is dramatically faster than parsing JSON again.

This proof compares three JSON fixture sizes against pre-generated DX .machine files. Generation time is excluded because the DX ecosystem creates .machine files ahead of runtime. Each path reads the same payload shape and checksum-verifies the same data.

case JSON size .machine size JSON parse JSON read + parse .machine validated .machine hot mmap
small 85,969 B 41,824 B 305.693us 416.907us 28.001us 1.334us
medium 2,751,387 B 1,336,240 B 8.667ms 10.461ms 675.906us 55.974us
large 22,024,633 B 10,690,480 B 98.201ms 129.378ms 5.538ms 499.677us
case validated .machine vs JSON parse hot mmap .machine vs JSON parse hot mmap .machine vs JSON read + parse
small 10.92x faster 229.11x faster 312.47x faster
medium 12.82x faster 154.85x faster 186.90x faster
large 17.73x faster 196.53x faster 258.92x faster

Brutal truth: JSON is excellent as an interchange format, but it is the wrong runtime read model for already-known data. DX .machine turns JSON-derived data into a compact binary read model that can be mmaped and accessed without reparsing. That is the unlock.

Machine Format Performance

DX-Machine uses pure RKYV - identical performance:

  • Single serialize: ~48-51ns (RKYV: ~48ns, within 6% variance)
  • Batch 100: ~7.5µs (RKYV: ~7.9µs, actually 5% faster)
  • Zero-copy deserialization (identical to RKYV)
  • Production-ready and battle-tested Implementation: Zero-overhead wrapper with #[inline(always)] that compiles to identical machine code as RKYV.

Three Formats

DX Serializer uses a revolutionary 3-format system:

Human Format (.sr files on disk)

Beautiful, readable format that developers edit directly:

  • TOML/INI-like syntax with aligned = at column 28
  • Lives on real disk where you work (e.g., dx, package.sr)
  • Easy to read, write, and version control
  • This is the source of truth - you edit these files

LLM Format (.llm in .dx/serializer/)

Token-optimized format for AI context windows:

  • 52-73% token savings vs JSON
  • Compact notation with schema headers
  • Auto-generated in .dx/serializer/*.llm folder
  • Never edit manually - regenerated from human format

Machine Format (.machine in .dx/serializer/)

Pure RKYV binary format for maximum performance:

  • Zero-copy deserialization
  • ~48-51ns serialize time
  • Auto-generated in .dx/serializer/*.machine folder
  • Identical to RKYV wire format

Architecture: Human format files live on disk. When you save a dx file (or any file with DX serializer syntax), the extension automatically generates the .llm and .machine versions in the .dx/serializer/ folder. The .dx/ folder is gitignored as it contains generated files.

Note: DX-Machine IS RKYV. We use RKYV's wire format directly with no modifications.

Usage

# Human format files live on disk (you edit these)
# Example: dx, package.sr

# When you save a file, the extension auto-generates:
# .dx/serializer/dx.llm       (LLM-optimized, 52-73% token savings)
# .dx/serializer/dx.machine   (binary, zero-copy)

# CLI usage (if needed manually):
dx serializer dx          # Process single file
dx serializer .                  # Process directory recursively
dx serializer src/               # Process specific directory

Workflow:

  1. Edit human format files on disk (e.g., dx, package.sr)
  2. Save the file
  3. Extension automatically generates .llm and .machine in .dx/serializer/
  4. .dx/ folder is gitignored (contains generated files)
  5. Only commit the human format files

LLM Format

author=essensefromexistence
version=0.0.1
name=dx
description="Orchestrate dont just own your code"
title="Enhanced Developing Experience"
driven(path=@/driven)
editors(default=neovim items=[neovim zed vscode cursor antigravity replit "firebase studio"])
forge(repository="https://dx.vercel.app/essensefromexistence/dx" container=none pipeline=none tasks=none tools=[cli docs examples packages scripts style tests])
dependencies[name version](
dx-package-1 0.0.1
dx-package-2 0.0.1
)
js.dependencies(next=16.0.1 react=19.0.1)

Syntax Features

  • Scalars: key=value (no spaces around =)
  • Objects: name(key=val key2=val2) (parentheses, space-separated pairs)
  • Arrays: key=[item1 item2 item3] (square brackets, space-separated)
  • Tables: name[col1 col2 col3](row1_val1 row1_val2 row1_val3\nrow2_val1...) (headers in brackets, rows in parentheses)
  • Strings: Use quotes "..." for multi-word strings
  • Booleans: true/false
  • Null: null

Objects

config(host=localhost port=5432 debug=true)
server(url="https://api.example.com" timeout=30)

Arrays

tags=[rust performance serialization]
editors=[neovim zed "firebase studio"]

Tables

users[id name email](
1 Alice alice@ex.com
2 Bob bob@ex.com
)

Nested Sections

Use dot notation:

js.dependencies[react=19.0.1,next=16.0.1]
i18n.locales[path=@/locales,default=en-US]

Complete Example

name=dx
version=0.0.1
title="Enhanced Developing Experience"
workspace(paths=[@/www @/backend])
editors(items=[neovim zed vscode] default=neovim)
forge(repository="https://github.com/user/repo" tools=[cli docs tests])
js.dependencies(react=19.0.1 next=16.0.1)

Human Format Example

author = essensefromexistence
version = 0.0.1
name = dx
description = Orchestrate dont just own your code
title = Enhanced Developing Experience

[driven]
path = @/driven

[editors]
default = neovim
items:
- neovim
- zed
- vscode
- cursor
- antigravity
- replit
- firebase-studio

[workspace]
paths:
- @/www
- @/backend

[dependencies:1]
name = dx-package-1
version = 0.0.1

[dependencies:2]
name = dx-package-2
version = 0.0.1

Syntax Features

  • Scalars: key = value (spaces around = for readability)
  • Sections: [section] headers (TOML/INI-like)
  • Arrays: key: followed by - item lines
  • Nested Sections: [section.subsection]
  • Strings: Use quotes for multi-word strings: title = "My Title"
  • Alignment: Keys padded with spaces for column alignment (typically at column 28)

Scalars

key = value
title = "My Title"

Arrays

key:
- item1
- item2
- item3

Sections

[section]
key = value

[section.subsection]
key = value

Complete Example

name = dx
version = 0.0.1
title = "Enhanced Developing Experience"

[workspace]
paths:
- @/www
- @/backend

[editors]
items:
- neovim
- zed
- vscode
default = neovim

[forge]
repository = https://github.com/user/repo
tools:
- cli
- docs
- tests

[js.dependencies]
react = 19.0.1
next = 16.0.1

Format Locations

Architecture Overview:

  • Human format - Lives on real disk, you edit these files directly

    • Examples: dx, package.sr
    • Source of truth, version controlled in git
    • TOML/INI-like syntax with aligned = at column 28
  • LLM format (.llm) - Auto-generated in .dx/serializer/ folder

    • Never edit manually
    • Regenerated automatically when human format changes
    • 52-73% token savings vs JSON
  • Machine format (.machine) - Auto-generated in .dx/serializer/ folder

    • Binary format (pure RKYV)
    • Zero-copy deserialization
    • ~48-51ns serialize time

The .dx/ folder is gitignored as it contains generated files. Only commit human format files.

Machine Format (RKYV)

DX-Machine IS RKYV - we use RKYV directly:

  • Pure RKYV wire format (no modifications)
  • Zero-overhead wrapper with #[inline(always)]
  • Identical performance: ~48-51ns single, ~7.5µs batch 100
  • Zero-copy deserialization
  • Production-ready

The machine format is a binary representation using RKYV's archived data structures. It provides the fastest serialization/deserialization with zero-copy access to data.

use serializer::machine::{serialize, deserialize};
// Serialize (calls rkyv::to_bytes directly)
let bytes = serialize(&data)?;
// Deserialize (calls rkyv::access_unchecked directly)
let archived = unsafe { deserialize::<MyType>(&bytes) };

Key Characteristics

  • Binary Format: Pure binary data, not human-readable
  • Zero-Copy: Direct memory access without copying data
  • Performance: Sub-nanosecond access times
  • Safety: Uses RKYV's compile-time validation
  • Compatibility: Identical to RKYV's wire format

Machine Format Compression

DX Machine format supports optional compression using LZ4 and ZSTD algorithms to reduce wire size while maintaining fast decompression.

Compression Algorithms

LZ4 Compression (Default)

  • Speed: Extremely fast compression/decompression
  • Ratio: Good compression for structured data
  • Use Case: Network transfer, storage where speed is critical
  • Pure Rust: No C dependencies (lz4_flex crate)

ZSTD Compression

  • Speed: Fast compression, very fast decompression
  • Ratio: Excellent compression ratios (better than LZ4)
  • Use Case: Maximum size reduction, archival storage
  • Levels: 1 (fast), 3 (balanced), 19 (maximum compression)

Usage

use serializer::machine::compress::{DxCompressed, CompressionLevel};

// Compress data with LZ4 (fast, default)
let compressed = DxCompressed::compress(b"your binary data here");

// Compress with specific level
let compressed = DxCompressed::compress_level(b"data", CompressionLevel::High);

// Get compression stats
println!("Original: {} bytes", compressed.original_size());
println!("Compressed: {} bytes", compressed.compressed_size());
println!("Ratio: {:.2%}", compressed.ratio());
println!("Space saved: {:.1}%", compressed.savings() * 100.0);

// Decompress (lazy - first access triggers decompression)
let data = compressed.decompress()?;

// Check if already decompressed (cached)
if compressed.is_cached() {
    println!("Data is cached in memory");
}

Streaming Compression

For large datasets, use streaming compression to process data in chunks:

use serializer::machine::compress::StreamCompressor;

// Create streaming compressor (64KB chunks)
let mut compressor = StreamCompressor::default_chunk();

// Write data in chunks
compressor.write(&large_data_chunk_1);
compressor.write(&large_data_chunk_2);

// Finish and get compressed chunks
let chunks = compressor.finish();

// Each chunk is individually compressed
for chunk in chunks {
    println!("Chunk: {}{} bytes", 
        chunk.original_size(), chunk.compressed_size());
}

Cargo Features

Enable compression features in your Cargo.toml:

[dependencies]
dx-serializer = { version = "0.1", features = ["compression"] }

# Or enable specific algorithms:
dx-serializer = { version = "0.1", features = ["compression-lz4", "compression-zstd"] }

Performance Characteristics

Algorithm Compression Speed Decompression Speed Ratio Use Case
LZ4 ~500 MB/s ~2000 MB/s 50-70% Network, real-time
ZSTD-1 ~300 MB/s ~1000 MB/s 60-80% Fast compression
ZSTD-3 ~100 MB/s ~800 MB/s 70-85% Balanced
ZSTD-19 ~10 MB/s ~500 MB/s 75-90% Maximum compression

Wire Format

Compressed data includes size prepending for safe decompression:

[original_size: u32][compressed_data...]

This allows safe decompression without knowing the original size in advance.

Conversion Rules

LLM → Human

  • Objects name(key=val) become [name] sections with key-value pairs
  • Arrays key=[item1 item2] become key: followed by - item lines
  • Keys are padded for alignment
  • Nested sections use dot notation

Human → LLM

  • [section] headers with key-value pairs become section(key=val)
  • key: followed by - item lines become key=[item1 item2]
  • All whitespace padding is removed
  • Numbered sections combine into tables

Why DX Beats TOON

  • No indentation - TOON requires 2 spaces per level
  • Inline objects - section:count[key=value] vs nested YAML
  • Space-separated arrays - No commas needed
  • Tabular data - name:count(schema)[rows] for structured data
  • Prefix elimination - @prefix removes repeated prefixes

Quick Start

use serializer::{json_to_dx, dx_to_json};
let json = r#"{"name": "app", "version": "1.0"}"#;
let dx = json_to_dx(json)?;

Features

[dependencies]
dx-serializer = { version = "0.1", features = ["tiktoken"] }

+--------------+----------------+ | Feature | Description | +==============+================+ | converters | JSON/YAML/TOML | +--------------+----------------+

Documentation

License

MIT / Apache-2.0