eavec 0.1.0

SIMD-accelerated vector similarity search. 3-4x faster than FAISS at dim=384.

eavec

SIMD-accelerated vector similarity search. Pre-compiled kernels with a NumPy API.

pip install eavec

Quick start

import numpy as np
from eavec import batch_cosine, top_k

db = np.random.rand(100_000, 384).astype(np.float32)
query = np.random.rand(384).astype(np.float32)

scores = batch_cosine(query, db)
values, indices = top_k(scores, k=10)

Why eavec?

• 3-4x faster than FAISS at dim=384 (the most common embedding dimension)
• 8-27x faster than NumPy depending on metric and dimension
• Zero dependencies beyond NumPy. No Rust, no LLVM, no compilation step.
• Multi-threaded batch operations release the GIL for true parallelism.

API

Single-pair operations

Function	Signature	Returns
dot(a, b)	two 1D float32 arrays	float
norm(a)	one 1D float32 array	float
cosine(a, b)	two 1D float32 arrays	float
l2_squared(a, b)	two 1D float32 arrays	float

Batch operations

Function	Signature	Returns
batch_dot(query, db)	query: 1D(dim,), db: 2D(n,dim)	1D float32(n,)
batch_cosine(query, db)	query: 1D(dim,), db: 2D(n,dim)	1D float32(n,)
batch_l2(query, db)	query: 1D(dim,), db: 2D(n,dim)	1D float32(n,)
normalize(db)	db: 2D(n,dim)	2D float32(n,dim)
threshold_filter(scores, threshold)	scores: 1D float32	1D int32 indices
top_k(scores, k)	scores: 1D float32, k: int	(values, indices)

Multi-threaded

Function	Signature	Returns
batch_dot_mt(query, db, num_threads=None)	same + threads	1D float32(n,)
batch_cosine_mt(query, db, num_threads=None)	same + threads	1D float32(n,)
batch_l2_mt(query, db, num_threads=None)	same + threads	1D float32(n,)

Platforms

• Linux x86_64 (AVX2)
• Linux aarch64 (NEON)
• Windows x86_64 (AVX2)

How it works

eavec ships pre-compiled SIMD kernels built with Ea, a SIMD kernel compiler. The kernels use dual-accumulator FMA with f32x8 vectors, restrict pointers for no-alias optimization, and explicit prefetch hints.

License

Apache-2.0