PuppyDB 🐶

• Flat files + mmap for fast vector storage
• LMDB for metadata and offsets
• HNSW (Hierarchical Navigable Small World graphs) for approximate nearest neighbor search
• Simple and hackable architecture

Fetch your vectors fast — just like a good puppy! 🐾

Project Status

🚧 In Progress
This is an experimental learning project to explore how vector databases work under the hood.
Major components are being built step by step.

Features

• ✅ Persistent vector + metadata storage
• ✅ Brute-force KNN search
• ✅ Approximate nearest neighbor search (HNSW-based)
• 🛠️ Lightweight and fun to use
• 🛠️ Comprehensive test suite (pytest)

Project Goals

• Learn the internals of building a vector database
• Implement a custom HNSW search
• Keep the architecture simple and hackable
• Provide clean examples and tests

Installation

You can install PuppyDB from PyPI:

pip install pup-py-db==0.1.0

Usage

from puppydb.core import PuppyDB
import numpy as np

# Initialize PuppyDB
db = PuppyDB("puppydb/data/vectors.bin", "puppydb/data/metadata_store")

# Insert vector
vector_id = "vec_001"
vector = np.random.randn(512).astype('float32')
metadata = {"user": "alice"}

db.insert_vector(vector_id, vector, metadata)

# Retrieve vector
retrieved = db.get_vector(vector_id)
if retrieved:
    retrieved_vector, retrieved_metadata = retrieved
    print("Vector metadata:", retrieved_metadata)

# Delete vector
db.delete_vector(vector_id)

# Truncate entire DB
db.truncate()

# Close DB
db.close()

License

MIT License — use and modify freely.