vctrs 0.1.5

A fast embedded vector database. Rust core, Python API. 4-10x faster than ChromaDB.

vctrs

A fast embedded vector database. Rust core, Python API.

Built for apps that need vector search without running a separate server — RAG pipelines, desktop apps, CLI tools, notebooks.

Install

pip install vctrs

Usage

import numpy as np
from vctrs import Database

# Create or open a database
db = Database("./mydb", dim=384, metric="cosine")

# Add vectors (accepts lists or numpy arrays)
db.add("doc1", np.random.rand(384).astype(np.float32), {"title": "hello"})

# Batch insert (much faster)
ids = [f"doc{i}" for i in range(10000)]
vectors = np.random.rand(10000, 384).astype(np.float32)
db.add_many(ids, vectors)

# Search
results = db.search(query_vector, k=10)
for id, distance, metadata in results:
    print(f"{id}: {distance:.4f}")

# Update
db.update("doc1", vector=new_vector, metadata={"title": "updated"})

# Delete
db.delete("doc1")

# Check membership
"doc1" in db
len(db)

# Persist to disk (graph structure saved — instant reload)
db.save()

Metrics

• "cosine" (default) — cosine distance
• "euclidean" / "l2" — squared L2 distance
• "dot" / "dot_product" — negative dot product

Tuning search

# ef_search controls recall vs speed. Higher = better recall, slower.
results = db.search(query, k=10, ef_search=300)

Benchmarks

100,000 vectors, 384 dimensions, Apple M-series:

Speed

Operation	vctrs	ChromaDB	numpy brute-force
Insert 100k	39s	70s	—
Search k=10	0.89ms	2.04ms	3.44ms
Search k=100	0.72ms	2.49ms	—
Load from disk	323ms	—	—
Get by id	<0.01ms	—	—

3-4x faster than ChromaDB on search, 4x faster than numpy brute-force.

Recall (search quality)

Measured against brute-force ground truth at 10k vectors (higher is better):

k	vctrs	ChromaDB
1	92%	76%
10	91%	76%
50	85%	68%

vctrs is both faster and more accurate than ChromaDB out of the box.

How it works

• HNSW index for O(log n) approximate nearest neighbor search
• SimSIMD for hardware-accelerated distance computation (ARM NEON, x86 AVX2/512)
• Rayon for parallel index construction
• Graph serialization — saves the full HNSW structure so loading is instant (no rebuild)
• PyO3 + maturin for zero-copy Python bindings with numpy support

Building from source

pip install maturin
git clone https://github.com/yang-29/vctrs.git
cd vctrs
python -m venv .venv && source .venv/bin/activate
pip install numpy maturin
maturin develop --release

License

MIT