polarisdb 0.1.1

High-performance embedded vector database

PolarisDB Python Bindings

PolarisDB is a high-performance, embedded vector database written in Rust. This package provides Python bindings for efficient local vector search and storage.

Features

• 🚀 Fast: Built on Rust with SIMD optimizations.
• 💾 Embedded: Runs locally without a separate server process.
• 📦 Simple: Easy-to-use Python API for managing collections and indexes.
• 🔍 Search: Supports Euclidean, Cosine, and Dot Product distance metrics.
• 🛡️ Durable: WAL-based persistence for data safety (Collection API).

Installation

pip install polarisdb

Quick Start

In-Memory Index (Brute Force)

Perfect for small datasets or exact search requirements.

import numpy as np
from polarisdb import Index

# Create an index with Cosine similarity and 128 dimensions
index = Index("cosine", 128)

# Insert vectors (ID, Vector)
# Vectors can be lists or numpy arrays
v1 = np.random.rand(128).astype(np.float32)
index.insert(1, v1)

v2 = np.random.rand(128).astype(np.float32)
index.insert(2, v2)

# Search for nearest neighbors
query = np.random.rand(128).astype(np.float32)
results = index.search(query, k=5)

for id, distance in results:
    print(f"Found ID: {id}, Distance: {distance}")

Persistent Collection

Store vectors on disk with crash recovery.

from polarisdb import Collection

# Open or create a collection at the specified path
collection = Collection.open_or_create("./my_collection", 128, "euclidean")

# Insert vectors
v1 = [1.0] * 128
collection.insert(1, v1)

# Persist changes to disk (WAL checkpoint)
collection.flush()

# Search
results = collection.search(v1, k=1)
print(results)

API Reference

Index(metric: str, dimension: int)

Creates an in-memory brute-force index.

• metric: "cosine", "euclidean", or "dot".
• dimension: Dimension of vectors (e.g., 128, 768).

Collection.open_or_create(path: str, dimension: int, metric: str)

Opens an existing collection or creates a new one.

• path: Directory path for storage.
• dimension: Vector dimension.
• metric: Distance metric.

License

MIT