🐹 NoPokeDB

A tiny, disk-backed vector database built on hnswlib (for approximate nearest neighbor search) and SQLite (for metadata).
Simple, lightweight, and crash-safe — good for prototypes, side projects, or when you don’t want a heavyweight vector DB service.

✨ Features

• Durable: HNSW index + SQLite metadata, persisted on disk
• Crash safety: write-ahead oplog with replay
• Fast: batch inserts, bulk metadata fetch, auto-resize index
• Flexible metrics: cosine / L2 / inner product
• CRUD: add, add_many, get, delete, upsert
• Consistent scoring: higher = better for all metrics (with raw distance included)

📦 Installation

pip install nopokedb

🚀 Quickstart

import numpy as np
from nopokedb import NoPokeDB

# Create (or load existing) DB with 128-d vectors
db = NoPokeDB(dim=128, max_elements=10_000, path="./vdb_data", space="cosine")

# Insert one vector
vec = np.random.rand(128).astype(np.float32)
vid = db.add(vec, metadata={"name": "foo"})

# Insert many at once (faster)
V = np.random.rand(5, 128).astype(np.float32)
metas = [{"i": i} for i in range(len(V))]
ids = db.add_many(V, metas)

# Query nearest neighbors
q = np.random.rand(128).astype(np.float32)
hits = db.query(q, k=3)
for h in hits:
    print(h["id"], h["score"], h["metadata"])

# Get / update / delete
print(db.get(vid))
db.upsert(vid, metadata={"name": "bar"})
db.delete(ids[0])

# Persist & close
db.save()
db.close()

Example query result:

{'id': 0, 'metadata': {'name': 'foo'}, 'score': 0.991, 'distance': 0.009}

🛠 Development

Bump version

The Makefile automates version bumps in pyproject.toml and src/nopokedb/__init__.py:

make bump V=0.2.0

That will:

• update both files
• commit with message Bump to v0.2.0
• tag v0.2.0
• push tag + main branch

Publish to PyPI

1. Set your API key once:

   export PYPI_API_KEY=your-pypi-token
   

2. Build + upload:

   make publish
   

📜 License

MIT