ETD

• seismic, SIMD, integer compression, dot product

Many state-of-the-art systems for top-k similarity search on sparse embeddings, such as Seismic, rely on a forward index to compute fast dot products between document and query vectors generated by Learned Sparse Retrieval (LSR) models. The goal of this thesis is to compress this index to reduce its memory footprint without compromising retrieval latency. To achieve this, we propose specialized SIMD-integrated scoring codecs, namely DotPacking8 and DotPackingDp, which directly fuse decompression with dot product evaluation to overcome the memory bandwidth bottleneck. Furthermore, we design a clustered reference-based approach to exploit semantic redundancies across embeddings, significantly improving compression density. Our methodologies were extensively benchmarked against various state-of-the-art integer compressors, providing a comprehensive evaluation of the space-time trade-off between storage efficiency and query latency.