🔥 ignite-vec-db

A lightweight and educational vector database written in TypeScript, powered by ONNX Runtime, RocksDB, and modern indexing algorithms like HNSW.

🚀 Project Goals

This project is built for learning purposes. It helps developers understand how vector databases work by providing a minimal but extensible implementation using familiar tools from the Node.js ecosystem.

✨ Features

• Embedding text with ONNX models (e.g., all-MiniLM-L6-v2)
• Persistent storage using RocksDB or SQLite
• Vector similarity search (brute-force, HNSW)
• Metadata filtering support
• CLI and programmatic API
• Type-safe validation with Zod
• Optional GPU acceleration via ONNX Runtime (CUDA)

🧱 Stack

Component	Tech
Language	TypeScript
Embedding Engine	ONNX Runtime (onnxruntime-node)
Storage Engine	RocksDB (node-rocksdb)
Indexing	Brute-force, hnswlib-node
CLI Interface	commander
Validation	zod
Testing	vitest

📦 Installation

pnpm install

Add runtime dependencies:

pnpm add onnxruntime-node hnswlib-node rocksdb zod commander

Add dev dependencies:

pnpm add -D typescript tsx vitest @types/node

Optional: for GPU support (requires CUDA 11.8+):

pnpm add onnxruntime-node-gpu

🗂 Project Structure

ignite-vec-db/
├── src/
│   ├── cli/          # Command-line interface (commander)
│   ├── db/           # Persistence (RocksDB / SQLite)
│   ├── index/        # Indexing algorithms
│   ├── models/       # ONNX Runtime embedder
│   ├── core/         # Vector math, engine core
│   └── utils/
├── onnx/             # ONNX models
├── tests/            # Vitest-based test suite
├── scripts/          # Helpers for data ingestion, etc.
└── README.md

🧪 Examples

Add a document

pnpm tsx src/cli/index.ts add "Artificial intelligence is transforming the world."

Search for similar documents

pnpm tsx src/cli/index.ts search "Machine learning innovations"

🛠 Future Features

• Web UI with React
• WebAssembly support for browser-side DB
• Advanced filtering (facets, tags)
• Vector quantization (PQ / IVFPQ)
• Embedding model loading from HuggingFace

👨‍🏫 Educational Focus

This project is not meant to compete with Pinecone, Weaviate, or FAISS — it's a clean-room reimplementation to demystify vector databases and give developers hands-on experience with embeddings, indexing, and similarity search at a low level.

---

Made with ❤️ for learning.