OpenSearch Neural Sparse Encoding Doc v3 GTE ONNX

This repository contains an ONNX export of opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte, converted by Qdrant for inference with FastEmbed.

Model Description

This is a learned sparse retrieval model.

It encodes documents into 30,522-dimensional sparse vectors. Queries are encoded using a tokenizer and a weight lookup table. A non-zero dimension represents the corresponding token in the vocabulary, and its value represents the importance of that token.

The similarity score is the inner product of the query and document sparse vectors.

Usage

This ONNX model is designed for use with FastEmbed.

from fastembed import SparseTextEmbedding

model = SparseTextEmbedding(
    model_name="opensearch-project/opensearch-neural-sparse-encoding-doc-v3-gte"
)

queries = ["What's the weather in New York now?"]
documents = ["Currently New York is rainy."]

query_embeddings = list(model.query_embed(queries))
document_embeddings = list(model.embed(documents))

Use query_embed() for queries and embed() for documents.

For the original PyTorch model and additional usage examples, refer to the original model repository.

Performance

The original model card reports an average NDCG@10 of 0.546 and average FLOPS of 1.7 on the evaluated subset of BEIR.

For detailed benchmark results, refer to the original model card.

License

This project is licensed under the Apache v2.0 License.

Copyright

Copyright OpenSearch Contributors. See NOTICE for details.

Acknowledgments

The original model was developed by the OpenSearch project. The ONNX conversion was performed by Qdrant for use with FastEmbed.