light-embed 1.0.3

Fast and Lightweight Text Embedding

LightEmbed

LightEmbed is a light-weight, fast, and efficient tool for generating sentence embeddings. It does not rely on heavy dependencies like PyTorch and Transformers, making it suitable for environments with limited resources.

Benefits

1. Light-weight

• Minimal Dependencies: LightEmbed does not depend on PyTorch and Transformers.
• Low Resource Requirements: Operates smoothly with minimal specs: 1GB RAM, 1 CPU, and no GPU required.

2. Fast (as light)

• ONNX Runtime: Utilizes the ONNX runtime, which is significantly faster compared to Sentence Transformers that use PyTorch.

3. Consistent with Sentence Transformers

• Consistency: Incorporates all modules from a Sentence Transformer model, including normalization and pooling.
• Accuracy: Produces embedding vectors identical to those from Sentence Transformers.

4. Supports models not managed by LightEmbed

LightEmbed can work with any Hugging Face repository, even those not hosted on Hugging Face ONNX models, as long as ONNX files are available.

5. Local Model Support

LightEmbed can load models from the local file system, enabling faster loading times and functionality in environments without internet access, such as AWS Lambda or EC2 instances in private subnets.

Installation

pip install -U light-embed

Usage

Then you can specify the original model name like this:

from light_embed import TextEmbedding
sentences = ["This is an example sentence", "Each sentence is converted"]

model = TextEmbedding(model_name_or_path='sentence-transformers/all-MiniLM-L6-v2')
embeddings = model.encode(sentences)
print(embeddings)

or, alternatively, you can specify the onnx model name like this:

from light_embed import TextEmbedding
sentences = ["This is an example sentence", "Each sentence is converted"]

model = TextEmbedding(model_name_or_path='onnx-models/all-MiniLM-L6-v2-onnx')
embeddings = model.encode(sentences)
print(embeddings)

Using a Non-Managed Model: To use a model from its original repository without relying on Hugging Face ONNX models, simply specify the model name and provide the model_config, assuming the original repository includes ONNX files.

from light_embed import TextEmbedding
sentences = ["This is an example sentence", "Each sentence is converted"]

model_config = {
    "model_file": "onnx/model.onnx",
    "pooling_config_path": "1_Pooling",
    "normalize": False
}
model = TextEmbedding(
    model_name_or_path='sentence-transformers/all-MiniLM-L6-v2',
    model_config=model_config
)
embeddings = model.encode(sentences)
print(embeddings)

Using a Local Model: To use a local model, specify the path to the model's folder and provide the model_config.

from light_embed import TextEmbedding
sentences = ["This is an example sentence", "Each sentence is converted"]

model_config = {
    "model_file": "onnx/model.onnx",
    "pooling_config_path": "1_Pooling",
    "normalize": False
}
model = TextEmbedding(
    model_name_or_path='/path/to/the/local/model/all-MiniLM-L6-v2-onnx',
    model_config=model_config
)
embeddings = model.encode(sentences)
print(embeddings)

The model_config is a dictionary that provides details about the model, such as the location of the ONNX file and whether pooling or normalization is needed. Pooling is required if it hasn't been incorporated into the ONNX file itself.

model_config = {
    "model_file": "relative path to the onnx file, e.g., model.onnx, or onnx/model.onnx",
    "pooling_config_path": "relative path to the pooling config folder, e.g., 1_Pooling",
    "normalize": True/False
}

If the pooling has been incorporated into the ONNX file, you can ignore the "pooling_config_path". Similarly, if normalization is already included in the ONNX file, you can omit the "normalize" entry.

Citing & Authors

Binh Nguyen / [e-mail]