padmadb 0.1.3

A blazing-fast, lightweight in-memory vector database built in pure Python.

PadmaDB 🪷

A blazing-fast, lightweight in-memory vector database built in pure Python and NumPy.

Pinecone, Milvus, and Weaviate are incredible for enterprise-scale billions of vectors. But when you are building local AI apps, testing RAG pipelines, or running lightweight microservices, spinning up a massive cloud infrastructure is overkill.

PadmaDB solves this by providing a hyper-fast, zero-dependency (beyond NumPy/FastAPI) vector search engine that runs locally and persists to disk.

---

🚀 Features

• Pure Python/NumPy Engine: No C++ compilers, Docker containers, or cloud APIs required.
• Exact Nearest Neighbor (k-NN): Guarantees 100% recall using optimized float32 Cosine Similarity matrix multiplications.
• RESTful Architecture: Ships with a built-in FastAPI server, allowing it to act as an independent microservice database.
• Local Persistence: Automatically serializes vector state to your local disk.

---

📦 Installation

pip install padmadb

---

⚡ Quickstart

1. Start the Database Server

You can launch the server locally in just a few lines of code:

from padmadb.server import run_server

if __name__ == "__main__":
    # Runs the database engine on [http://0.0.0.0:8000](http://0.0.0.0:8000)
    run_server()

2. Connect and Query (Client SDK)

In a separate Python script or Jupyter Notebook, connect to your local PadmaDB instance to insert and search vectors.

from padmadb import PadmaClient

# Initialize client
client = PadmaClient(host="http://localhost:8000")

# Insert a 384-dimensional vector (e.g., from all-MiniLM-L6-v2)
vector_data = [0.12, -0.45, 0.88] # Truncated for example; must be length 384
client.insert(
    vec_id="doc_1", 
    vector=vector_data, 
    metadata={"text": "Machine learning is fascinating."}
)

# Search for the Top-K nearest neighbors
results = client.search(vector=vector_data, top_k=1)
print(results)

---

🧠 Architecture & Roadmap

PadmaDB currently utilizes a brute-force exact search mechanism tailored for datasets up to 100,000 vectors, where in-memory matrix multiplication is effectively instantaneous.

Upcoming in v0.2.x:

• Implementation of threading.Lock() for concurrent write-safety.
• Transition from JSON persistence to memory-mapped .npy binary arrays for faster cold boots.
• HNSW (Hierarchical Navigable Small World) indexing for billion-scale querying with O(log N) search times.

---

🤝 Contributing

PadmaDB is fully open-source. We warmly welcome collaborators, researchers, and cloners!

Want to help build the ultimate local AI database? Clone the repo, fork it, break it, fix it, and submit a PR!