on2vec

A toolkit for generating vector embeddings from OWL ontologies using Graph Neural Networks (GNNs), with HuggingFace Sentence Transformers integration and MTEB benchmarking.

🚀 Quick Start

Installation

pip install on2vec

Create production-ready Sentence Transformers models with ontology knowledge in one command:

# Complete end-to-end workflow
on2vec hf biomedical.owl my-biomedical-model

Use like any sentence transformer:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('./hf_models/my-biomedical-model')
embeddings = model.encode(['heart disease', 'cardiovascular problems'])

📋 Table of Contents

• 🚀 Quick Start
• 📥 Installation
• 🤗 HuggingFace Integration
• 🧪 MTEB Benchmarking
• 💻 Core on2vec Usage
• 🏗️ Architecture
• 📚 Documentation

📥 Installation

From PyPI (Recommended)

# Basic installation
pip install on2vec

# With MTEB benchmarking support
pip install on2vec[benchmark]

# With all optional dependencies
pip install on2vec[all]

From Source

git clone <repository-url>
cd on2vec
pip install -e .

Dependencies

• Python >= 3.10
• PyTorch + torch-geometric
• owlready2, sentence-transformers
• polars, matplotlib, umap-learn

🤗 HuggingFace Integration

One-Command Model Creation

# Create complete model with auto-generated documentation
on2vec hf ontology.owl model-name

# With custom settings
on2vec hf ontology.owl model-name \
  --base-model all-mpnet-base-v2 \
  --fusion gated \
  --epochs 200

Step-by-Step Workflow

# 1. Train ontology embeddings
on2vec hf-train ontology.owl --output embeddings.parquet

# 2. Create HuggingFace model (auto-detects base model)
on2vec hf-create embeddings.parquet model-name

# 3. Test model functionality
on2vec hf-test ./hf_models/model-name

# 4. Inspect model details
on2vec inspect ./hf_models/model-name

Batch Processing

# Process multiple ontologies
on2vec hf-batch owl_files/ ./output \
  --base-models all-MiniLM-L6-v2 all-mpnet-base-v2 \
  --fusion-methods concat gated \
  --max-workers 4

Features

• ✅ Auto-generated model cards with comprehensive metadata
• ✅ Smart base model detection from embeddings
• ✅ Upload instructions and HuggingFace Hub preparation
• ✅ Domain detection and appropriate tagging
• ✅ Multiple fusion methods: concat, attention, gated, weighted_avg
• ✅ Batch processing for multiple ontologies

🧪 MTEB Benchmarking

Evaluate your models against the Massive Text Embedding Benchmark:

Quick Benchmark

# Fast evaluation on subset of tasks
on2vec benchmark ./hf_models/my-model --quick

# Focus on specific task types
on2vec benchmark ./hf_models/my-model --task-types STS Classification

# Full MTEB benchmark
on2vec benchmark ./hf_models/my-model

Compare Models

# Benchmark vanilla baseline
on2vec benchmark sentence-transformers/all-MiniLM-L6-v2 \
  --model-name vanilla-baseline --quick

# Compare ontology vs vanilla models
on2vec compare ./hf_models/my-model --detailed

Features

• ✅ Full MTEB integration with 58+ evaluation tasks
• ✅ Task filtering by category (STS, Classification, Clustering, etc.)
• ✅ Automated reporting with JSON summaries and markdown reports
• ✅ Resource management with configurable batch sizes
• ✅ Comparison tools for baseline evaluation

💻 Core on2vec Usage

Basic Training

# Train GCN model
on2vec train ontology.owl --output model.pt --model-type gcn --epochs 100

# Train with text features (for HuggingFace integration)
on2vec train ontology.owl --output embeddings.parquet --use-text-features

# Multi-relation models with all ObjectProperties
on2vec train ontology.owl --output model.pt --use-multi-relation --model-type rgcn

Generate Embeddings

# Generate embeddings from trained model
on2vec embed model.pt ontology.owl --output embeddings.parquet

Visualization

# Create UMAP visualization
on2vec visualize embeddings.parquet --output visualization.png

Python API

from sentence_transformers import SentenceTransformer
from on2vec import train_ontology_embeddings, embed_ontology_with_model

# Train model
result = train_ontology_embeddings(
    owl_file="ontology.owl",
    model_output="model.pt",
    model_type="gcn",
    hidden_dim=128,
    out_dim=64
)

# Generate embeddings
embeddings = embed_ontology_with_model(
    model_path="model.pt",
    owl_file="ontology.owl",
    output_file="embeddings.parquet"
)

# Use HuggingFace model
model = SentenceTransformer('./hf_models/my-model')
vectors = model.encode(['concept 1', 'concept 2'])

🏗️ Architecture

Core Components

• Graph Construction: Converts OWL ontologies to graph representations
• GNN Training: Supports GCN, GAT, RGCN, and heterogeneous architectures
• Text Integration: Combines structural and semantic features using sentence transformers
• Fusion Methods: Multiple approaches to combine text + structural embeddings
• HuggingFace Bridge: Creates sentence-transformers compatible models

Model Pipeline

OWL Ontology → Graph → GNN Training → Structural Embeddings
                                            ↓
Text Features → Sentence Transformer → Text Embeddings
                                            ↓
                              Fusion Layer → Final Model
                                            ↓
                              HuggingFace Model + Model Card

Supported Architectures

• GCN: Graph Convolutional Networks
• GAT: Graph Attention Networks
• RGCN: Relational GCN for multi-relation graphs
• Heterogeneous: Relation-specific layers with attention

📚 Documentation

• 📚 CLI Quick Reference - All commands and examples
• 📖 HuggingFace Integration - Complete workflow guide
• 🧪 MTEB Benchmarking - Evaluation framework
• 🧬 Project Instructions - Development guidelines

🎯 Key Features

• 🤗 HuggingFace Ready: One-command model creation with professional documentation
• 🧪 MTEB Integration: Comprehensive benchmarking against standard tasks
• 📊 Rich Metadata: Auto-generated model cards with complete technical details
• 🔧 Smart Automation: Auto-detects base models, domains, and configurations
• ⚡ Batch Processing: Handle multiple ontologies efficiently
• 🎨 Multiple Fusion Methods: Flexible combination of text and structural features
• 📈 Comprehensive Evaluation: Built-in comparison and testing tools

🚀 Example Workflow

# 1. Install on2vec
pip install on2vec[benchmark]

# 2. Create a model from biomedical ontology
on2vec hf EDAM.owl edam-biomedical

# 3. Quick benchmark evaluation
on2vec benchmark ./hf_models/edam-biomedical --quick

# 4. Compare with vanilla models
on2vec compare ./hf_models/edam-biomedical --detailed

# 5. Inspect model details
on2vec inspect ./hf_models/edam-biomedical

# 6. Upload to HuggingFace Hub (instructions auto-generated)
# See ./hf_models/edam-biomedical/UPLOAD_INSTRUCTIONS.md

The model is immediately usable as a drop-in replacement for any sentence-transformer, with the added benefit of ontological domain knowledge!

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

Citation

If you use on2vec in your research, please cite:

@software{on2vec2025,
  title={on2vec: Ontology Embeddings with Graph Neural Networks},
  author={David Steinberg},
  year={2025},
  url={https://github.com/david4096/on2vec}
}