libembedding

Forked from pacifio/libembedding. This fork adds Windows support (native DLL), PyPI packaging under the name libembedding-ng, local model loading, runtime introspection, similarity helpers, streaming, a multi-worker pool, an autotuner and automatic model selection. Imported as libembedding.

Fast ONNX-based text, image, and sparse embeddings for Python. Dynamic INT8 ONNX models are ~4x smaller, ~2x lighter in memory and ~1.7-2.0x faster than FP32 on CPU (measured; see Benchmarks below).

Built on a C/C++ backend using ONNX Runtime, exposed to Python via zero-overhead cffi bindings. The wheel bundles the compiled shared library (libembedding.so / .dylib / .dll) together with ONNX Runtime and libcurl, so no system ONNX Runtime install is required at runtime. Supports 44 text embedding models, 5 image models, 2 sparse models, and 4 rerankers with automatic model downloading from HuggingFace Hub.

Installation

pip install libembedding-ng

The PyPI package is named libembedding-ng (to avoid clashing with the original libembedding on PyPI); it is imported as libembedding. The wheel already bundles the native shared library and ONNX Runtime, so no separate system install is required.

Platforms: wheels exist for Windows x86-64, Linux x86-64 (manylinux), Linux aarch64 (manylinux_2_31: Raspberry Pi 3/4/5 under 64-bit Raspberry Pi OS, Ubuntu 20.04+, Graviton) and macOS arm64 (Apple Silicon, macOS 14 or later). The x86-64 wheels run on any x86-64 CPU, no AVX/AVX2 required: GGUF models load the ggml CPU library matching the machine at run time (libembedding.ggml_cpu_features() names it), and the aarch64 wheel does the same across eight ARM levels (armv8.0 to armv9.2). 32-bit ARM (armv7l) is not supported yet. On Linux, HTTPS downloads use the system CA bundle (SSL_CERT_FILE or CURL_CA_BUNDLE override it). The macOS wheel bundles its dependencies (ONNX Runtime, llama.cpp/ggml) under libembedding/.dylibs/ and uses the system libcurl.

Android: wheels for CPython 3.13 on Android 7.0+ (android_24_arm64_v8a, android_24_x86_64, PEP 738: BeeWare, Chaquopy). cffi and numpy come from Chaquopy's index: pip install --extra-index-url https://chaquo.com/pypi-13.1 libembedding-ng. Models download normally (static libcurl over mbedTLS, system certificate store). The x86_64 wheel is tested on an emulator in CI; the arm64 wheel has not been run on a device yet.

Requirements: ONNX Runtime must be installed on your system only if you build the package from source.

# macOS
brew install onnxruntime

# Ubuntu/Debian
apt install libonnxruntime-dev

# Or set ONNXRUNTIME_ROOT to your installation path

Quick Start

Text Embeddings

from libembedding import TextEmbedding

model = TextEmbedding("BAAI/bge-small-en-v1.5")
embeddings = model.embed(["Hello world", "How are you?"])

print(embeddings.shape)  # (2, 384)
print(embeddings.dtype)  # float32

Sparse Embeddings

from libembedding import SparseTextEmbedding

model = SparseTextEmbedding()
results = model.embed(["machine learning algorithms"])

for r in results:
    print(r.indices.shape, r.values.shape)

Image Embeddings

from libembedding import ImageEmbedding

model = ImageEmbedding()
embeddings = model.embed_files(["photo.jpg", "diagram.png"])

Reranking

from libembedding import Reranker

reranker = Reranker("BAAI/bge-reranker-base")
results = reranker.rerank(
    "What is deep learning?",
    [
        "Deep learning uses neural networks with many layers",
        "The weather is sunny today",
        "Neural networks are inspired by biological brains",
    ],
)
for r in results:
    print(f"doc[{r.index}] score={r.score:.4f}")

Model Discovery

import libembedding

for m in libembedding.list_text_models():
    print(f"{m.model_name:45} dim={m.dim:<5} {m.pooling}")

API Reference

Provider

TextEmbedding(
    model_name="BAAI/bge-small-en-v1.5",  # HuggingFace model name, repo code, or local dir path
    provider="cpu",  # "cpu", "cuda", "coreml", "directml", "tensorrt", "llamacpp"
    device_id=0,
    cache_dir=None,  # None = ~/.cache/libembedding
    max_length=0,  # 0 = model default
    threads=0,  # 0 = auto
    batch_size=256,  # internal batch size for embedding
    offline=False,  # True = use cache only, never download
    show_download_progress=True,
    dim=0,  # embedding dim for local models without config.json
    pooling="mean",  # "cls" or "mean" for local models
    num_threads=0,  # deprecated, use threads
)

SparseTextEmbedding

SparseTextEmbedding(model_name="prithvida/SPLADE_PP_en_v1", ...)

ImageEmbedding

ImageEmbedding(model_name="Qdrant/clip-ViT-B-32-vision", ...)

Reranker

Reranker(model_name="BAAI/bge-reranker-base", ...)

All classes support context managers (with TextEmbedding(...) as model:).

Similarity Functions

from libembedding import cosine_similarity, dot_product, euclidean_distance
import numpy as np

a = np.array([1.0, 2.0, 3.0], dtype=np.float32)
b = np.array([1.0, 2.0, 3.0], dtype=np.float32)

print(cosine_similarity(a, b))  # 1.0 (identical)
print(dot_product(a, b))  # 14.0
print(euclidean_distance(a, b))  # 0.0

Streaming Embeddings

Process large document sets without allocating a single result array:

from libembedding import TextEmbedding

with TextEmbedding("BAAI/bge-small-en-v1.5") as model:
    for embedding in model.embed_stream(["doc1", "doc2", ...], batch_size=32):
        # Each iteration yields a single (dim,) numpy array
        process(embedding)

Runtime Statistics

with TextEmbedding("BAAI/bge-small-en-v1.5") as model:
    model.embed(["text 1", "text 2", "text 3"])
    stats = model.stats()
    print(
        f"Embedded {stats.texts_embedded} texts "
        f"({stats.batches_run} batches), "
        f"avg latency {stats.avg_latency_ms:.2f}ms"
    )

Data Types

ModelDesc — runtime model descriptor (returned by info()):

Stats — runtime statistics (returned by stats()):

Local Model Loading

from libembedding import TextEmbedding

# Load from a local directory containing model.onnx + tokenizer.json (+ config.json)
model = TextEmbedding("/path/to/model_dir")

Available Models

44 text models including BGE, MiniLM, Nomic, E5, CLIP, Jina, GTE, Snowflake, ModernBERT (with quantized variants).

6 image models including CLIP ViT-B/32 (FP32 + INT8 quantized), ResNet-50, Unicom, Nomic Vision.

2 sparse models: SPLADE++, BGE-M3.

5 reranker models: BGE Reranker, Jina Reranker (including INT8 quantized variant).

GGUF/llama.cpp models: Run quantized GGUF models (Q4_K_M, Q8_0) via the llama.cpp backend.

Benchmarks

Comparison with fastembed — withdrawn

An earlier version of this file published a table claiming libembedding is 6.5-8.6x faster than fastembed, stated as "measured on Apple M-series". It has been removed. All four measurement harnesses hardcoded "platform": "macOS arm64" in their output, so no run could be traced to a machine, and the maintainer has no Mac to reproduce it on. The harnesses now report the real platform. See docs/archive/benchmarks/cross-implementation-macos-claim/ for the withdrawn figures and the reasoning.

Treat any cross-implementation speed claim about libembedding as unsubstantiated until it is re-measured on an identified machine.

Quantization — measured and reproducible

The quantization benchmark in the main README reports its full environment (CPU, ONNX Runtime version, execution provider) and reads the weight dtypes from the ONNX files, so its provenance is verifiable.

Unified Auto-Tuning

Auto-tune threads and batch_size (plus workers for text, max_tokens for the reranker) for any task type:

from libembedding import autotune, autotune_unified, clear_autotune_cache_unified

# Text embedding tune (QUICK by default; ~13 s cold on an i7-1065G7, indicative)
result = autotune("BAAI/bge-small-en-v1.5")

# Unified API for any task: "embedding", "reranking", "sparse", "image"
result = autotune_unified("embedding", "BAAI/bge-small-en-v1.5")

# Objective, custom workload and constraints (reranking: last text = the query)
result = autotune_unified(
    "reranking",
    "jinaai/jina-reranker-v1-turbo-en-quantized",
    objective="latency",              # "latency" | "throughput" | "balanced" | "memory"
    texts=[*documents, query],
    max_latency_ms=300.0,             # p95 budget for one tuning call (0 = none)
    min_tokens=128,                   # reranking only: floor for max_tokens
)
# Over budget? The best unconstrained config is returned: check result.p95_latency_ms.

clear_autotune_cache_unified("reranking")   # purge one task's cache

Results are cached per model, hardware, runtime, workload, objective, mode and constraints — a repeated call returns in milliseconds. For sparse models, top_k, min_weight and storage_format are not tuned (they change the output); the result reports the model defaults. See docs/performance_tuning.md for grids, methodology and measurements.

Error Handling

from libembedding import LembedError, LlamaError
from libembedding.exceptions import (
    ModelNotFoundError,
    OnnxRuntimeError,
    DownloadError,
)

try:
    model = TextEmbedding("nonexistent/model")
except ModelNotFoundError as e:
    print(f"Model not found: {e.detail}")
except LlamaError as e:
    print(f"llama.cpp error: {e.detail}")
except LembedError as e:
    print(f"Error [{e.status_code}]: {e.message}")

GGUF Models (llama.cpp)

Load quantized GGUF models:

from libembedding import TextEmbedding

# Check llama.cpp availability
if TextEmbedding.supports_llamacpp():
    model = TextEmbedding.from_gguf(
        repo="Xenova/all-MiniLM-L6-v2-GGUF",
        filename="all-MiniLM-L6-v2-Q4_K_M.gguf",
        provider="llamacpp",
    )
    embeddings = model.embed(["Hello world"])

Configuration

Building & Publishing

Prerequisites

pip install build twine

Build the shared library + wheel

cd python/

# Step 1: Build the C/C++ shared library and copy it into the package
./setup.sh --build-only

# Step 2: Build sdist and wheel
python -m build

This produces files in dist/ (versions follow the libembedding-ng package, e.g. 1.4.0):

dist/
  libembedding_ng-1.1.1.tar.gz                                # source distribution
  libembedding_ng-1.1.1-cp39-abi3-manylinux_2_28_x86_64.whl  # platform wheel (bundles libembedding.so + onnxruntime)

Upload to PyPI

# Upload to TestPyPI first to verify
twine upload --repository testpypi dist/*

# Install from TestPyPI to verify
pip install --index-url https://test.pypi.org/simple/ libembedding

# Upload to production PyPI
twine upload dist/*

One-liner (build + upload)

./setup.sh --build-only && python -m build && twine upload dist/*

Platform-specific wheels

The wheel is platform-tagged (e.g. manylinux, macosx, win_amd64) because it bundles the compiled shared library (libembedding.so / .dylib / .dll) along with ONNX Runtime and libcurl. python -m build produces a wheel for the current build platform. To build wheels for other platforms, build on that platform or use cibuildwheel:

# macOS (current arch)
./setup.sh --build-only
python -m build

# For other platforms, build on that platform or use cibuildwheel:
pip install cibuildwheel
cibuildwheel --platform linux   # builds manylinux wheels
cibuildwheel --platform macos   # builds macOS wheels

License

MIT