Research
Peer-reviewed papers from the tusKANNy team, covering approximate nearest neighbor search across dense, sparse, and multi-vector representations.
Franco Maria Nardini, Luca Rizzo, Cosimo Rulli, Rossano Venturini
Presents compact index layouts for learned sparse retrieval that reduce memory footprint while preserving query efficiency and accuracy.
Francesco Gargiulo, Rossano Venturini
Introduces OnPair, a dictionary-based compression algorithm for short strings that builds its dictionary in a single cache-friendly pass and compresses each string independently, matching BPE compression ratios with much faster training and fine-grained random access.
Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Describes how the kANNolo library and the Seismic index share a common foundation, showing how to combine them for efficient approximate retrieval over learned sparse representations.
Matteo Attimonelli, Alessandro De Bellis, Franco Maria Nardini, Claudio Pomo, Cosimo Rulli, Rossano Venturini, Tommaso Di Noia
Sparsifies pre-trained universal multimodal embeddings after training, enabling efficient retrieval with sparse indexes without retraining the embedding model.
Francesco Benocci, Leonard Dahlmann, Francesco Gargiulo, Tobias Heuer, Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Introduces lossless compression schemes for proximity graph indexes (e.g., HNSW) that shrink the index size without affecting search accuracy.
Angelo Savino, Rossano Venturini
Presents RISE, a Rust library covering the full inverted-index stack, from index construction and posting-list compression (Elias-Fano, PEF, Interpolative, Stream VByte) to WAND, MaxScore and Block-Max query processing, with speedups of up to 2x over the state of the art.
Andrew Yates, Carlos Lassance, Cosimo Rulli, Eugene Yang, Sean MacAvaney, Siddharth A.K. Singh, Thong Nguyen, Yibin Lei
Half-day tutorial on learned sparse retrieval: foundations (SPLADE-style models, training, evaluation, frameworks), efficient indexing, multilingual and multimodal extensions, and hybrid dense-sparse retrieval, with hands-on notebooks.
Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Proposes TACHIOM, a multivector retrieval system that uses Token-Aware Clustering (TAC) for accurate and scalable token clustering. By combining hierarchical indexing with a MaxSim-optimized Product Quantization layout, TACHIOM achieves up to 247x faster clustering than standard k-means and delivers up to 9.8x faster retrieval compared to state-of-the-art systems.
Read PDFThong Nguyen, Cosimo Rulli, Franco Maria Nardini, Rossano Venturini, Andrew Yates
Sparton is a Triton kernel for the Language Model head in Learned Sparse Retrieval models that fuses tiled matrix multiplication, ReLU, log1p, and max-reduction into a single GPU kernel, achieving up to 4.8x speedup and an order-of-magnitude reduction in peak memory usage compared to PyTorch baselines.
Read PDFSebastian Bruch, Martino Fontana, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Introduces DotVByte, a compression technique optimized for inner product computation that achieves significant space savings while maintaining sparse retrieval efficiency.
Read PDFSilvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Demonstrates that replacing token-level gatherer phases with learned sparse retrieval achieves over 24x speedup over state-of-the-art multivector retrieval systems.
Read PDFAndrew Yates, Carlos Lassance, Cosimo Rulli, Eugene Yang, Sean MacAvaney, Siddharth A.K. Singh, Thong Nguyen, Yibin Lei
Half-day tutorial on learned sparse retrieval: foundations (SPLADE-style models, training, evaluation, frameworks), efficient indexing, multilingual and multimodal extensions, and hybrid dense-sparse retrieval, with hands-on notebooks.
Leonardo Delfino, Domenico Erriquez, Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
A Rust-based ANN research library combining state-of-the-art indexing for dense and sparse vectors with vector quantization, designed for easy prototyping.
Read PDFSebastian Bruch, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini, Leonardo Venuta
Compares graph-based and inverted index-based sparse retrieval methods on the 138M-passage MS MARCO v2 dataset, uncovering scalability challenges and efficiency trade-offs.
Read PDFFranco Maria Nardini, Thong Nguyen, Cosimo Rulli, Rossano Venturini, Andrew Yates
Proposes Li-LSR, which replaces the query encoder with a fast lookup table by learning a static relevance score per token at training time, achieving state-of-the-art inference-free sparse retrieval and surpassing SPLADE-v3-Doc by 1 mRR@10 point on MsMarco and 1.8 nDCG@10 points on BEIR.
Read PDFAndrew Yates, Carlos Lassance, Cosimo Rulli, Eugene Yang, Sean MacAvaney, Siddharth A.K. Singh, Thong Nguyen, Yibin Lei
Half-day tutorial on learned sparse retrieval: foundations (SPLADE-style models, training, evaluation, frameworks), efficient indexing, multilingual and multimodal extensions, and hybrid dense-sparse retrieval, with hands-on notebooks.
Sebastian Bruch, Aditya Krishnan, Franco Maria Nardini
Studies query routing in clustering-based maximum inner product search and, following the principle of optimism in the face of uncertainty, proposes a framework that uses the moments of the inner-product distribution within each shard to estimate its maximum inner product and decide which shards to probe.
Silvio Martinico, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Investigates the feasibility of ANN search on a Raspberry Pi, a device two to three orders of magnitude cheaper than typical servers, through preliminary experiments with two state-of-the-art ANN libraries on dense and sparse data, identifying key challenges for ANN search in low-resource environments.
Read PDFSebastian Bruch, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Enhances Seismic with k-NN graph integration and a clustering hypothesis, achieving nearly 2.2x speedup over standard Seismic while maintaining accuracy.
Read PDFSebastian Bruch, Franco Maria Nardini, Cosimo Rulli, Rossano Venturini
Presents Seismic, a novel inverted index organization that enables fast retrieval over learned sparse embeddings, competitive with dense retrieval on BigANN benchmarks.
Read PDFFranco Maria Nardini, Cosimo Rulli, Rossano Venturini
Introduces EMVB, a multi-vector dense retrieval framework that speeds up PLAID-style centroid-based search with bit-vector pre-filtering, SIMD column-wise centroid computation, Product Quantization and per-document term filtering, achieving up to 2.8x faster queries and 1.8x lower memory footprint at no accuracy loss.
Sebastian Bruch, Franco Maria Nardini, Amir Ingber, Edo Liberty
Shows that IVF-style clustering, developed for dense vectors, is an efficient solution for sparse maximum inner product search, reframes IVF as a dynamic pruning technique for inverted indexes, and proposes a unified framework for hybrid dense-sparse vectors that is robust across query distributions.
Sebastian Bruch, Franco Maria Nardini, Amir Ingber, Edo Liberty
Introduces Sinnamon, an approximate top-k retrieval algorithm for streaming, real-valued sparse vectors from arbitrary distributions, with tunable trade-offs between memory, latency and accuracy, backed by a theoretical error analysis and an empirical evaluation across hardware platforms.