Skip to content

Repository files navigation

HDC Markdown Encoder & Reconstructor

HDC Python License Status screen-capture (1) webm

Transforms Markdown documents into hyperdimensional vectors and reconstructs them using dual HDC encoding and an optional LLM reconstruction.

How It Works

Dual HDC Pipeline: Documents split into semantic content and positional structure, encoded as 10,000-dimensional vectors, then reconstructed via HDC unbinding.

  1. Content Vector: Encodes which words exist
  2. Position Vector: Encodes which positions are used
  3. Pair Vectors: HDC binding creates word-position associations
  4. Reconstruction: HDC unbinding + LLM recovers original text

Key Innovation: Mathematical "handshake" between words and positions enables perfect order recovery.

Algorithm

HDC Binding: pair_vector = word_vector * position_vector
HDC Unbinding: recovered_word = document_vector * position_vector
Vector Bundling: content_vector = sum(word_vectors)
Storage: int8 (±1) for maximum compression more details in architecture_en.md

Key Features

  • Universal Dictionary: 20,000-word shared vocabulary
  • Scalable: Linear O(n) performance

Installation

git clone https://github.com/Garletz/HDC-Markdown-Encoder-Reconstructor.git

pip install -r requirements.txt && pip install -e .

CLI Usage

# Encode document
python cli.py --encode-dual "document.md" --config config.yaml

# Reconstruct from vectors
python cli.py --reconstruct-dual \
  --content-vector "encoded_vectors/encoded_X_content.npy" \
  --position-vector "encoded_vectors/encoded_X_position.npy" \
  -o "output.md"

Performance if word not repeated

Tokens Encoding Storage Reconstruction Accuracy
8 0.8s 240KB 1.2s 100%
16 1.1s 480KB 1.8s 100%
50+ 2.3s 1.5MB 3.1s 100%

Output Files

encoded_vectors/
├── encoded_N_content.npy    # Semantic information
├── encoded_N_position.npy   # Structural information
└── encoded_N_pairs.npy      # Word-position bindings

Limitations

  • Repeated words may cause position confusion (5-10 and more % cases)
  • Out-of-vocabulary tokens are skipped

  • ©OpenDataHive License (LICENSE MODIFIED)

Project Goal

  • Enable ultra-light document transfer via semantic vector compression.
    If sender and receiver share the same item memory (dictionary), the original text can be perfectly reconstructed from compact .npy vectors.
  • This approach aims to enable a wide range of future use cases...
  • Poneglyph ...

Currently experimental — concept in development.

📞 Contact

Get in touch with the OpenDataHive team:

Twitter Devpost WhatsApp


About

Transforms Markdown documents into hyperdimensional vectors and reconstructs them using dual HDC encoding and an universal dictionnary

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages