ESPN-Embedding-from-Storage-Retrieval-for-Multi-Vector-IR

ESPN: Embedding from Storage Pipelined Network. GDS implementation for multi-vector embedding retrieval and bindings.

Welcome to the ESPN repository. This repo is based on Nvidia's GPUDirect Storage architecture and allows embeddings to be retrieved directly from storage to GPU memory.

Citation

The full paper on ESPN can be found here: https://arxiv.org/abs/2312.05417

If you find our work helpful, please cite us:

@misc{shrestha2023espn,
      title={ESPN: Memory-Efficient Multi-Vector Information Retrieval}, 
      author={Susav Shrestha and Narasimha Reddy and Zongwang Li},
      year={2023},
      eprint={2312.05417},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}