@fmnardini

Research Director @ ISTI-CNR, Pisa, Italy. Information Retrieval, Machine/Deep Learning, Efficiency, Big Data.

 Pisa, Tuscany
hpc.isti.cnr.it/~nardini/
Joined March 2010
  • Tweets524
  • Following1.6K
  • Followers801
20Media
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoOct 16
Just received the award for the #SISAP2026 Indexing Challenge here in Brno! 🏆🇨🇿 Very proud of the tusKANNy team (@fmnardini, @cosimorulli1, @rventurini_). We took home both 1st and 2nd place in Task 3 (Learned Sparse Retrieval)! 🧵👇 #VectorSearch #InformationRetrieval
5
3
1
19
768
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoOct 13
The Vector Search landscape is full of complex methods. At tusKANNy, we like to simplify things through abstractions. Meet RerankIndex: a general framework for two-stage search, now available in kANNolo 0.12 (Python & Rust) 🧵👇 #VectorSearch #InformationRetrieval
5
8
1
29
1,125
Franco Maria Nardini retweeted
Sumit@_reachsumitOct 9
Compact and Efficient Indexes for Learned Sparse Retrieval @fmnardini et al. cut the memory of learned sparse retrieval indexes by using medoid documents as block representatives and compressing the forward index with bit-packing and 4-bit quantization 📝arxiv.org/abs/2610.12300

Compact and Efficient Indexes for Learned Sparse Retrieval

This paper investigates how to substantially reduce the memory footprint of learned sparse retrieval indexes without sacrificing the efficiency of state-of-the-art retrieval data structures....

arxiv.org
12
44
1,558
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoSep 29
Last week I published 9 multivector datasets. @Robro612 saw it and went all in, dropping 40+ new ones on HF! 🤯 The awesome-multivector-retrieval repo now tracks 50+ datasets ready to use. 17 corpora, 8 different encoders. github.com/TusKANNy/awesome-… #LateInteraction

GitHub - tusKANNy/awesome-multivector-retrieval: An extensive and commented list of resources on...

An extensive and commented list of resources on Late-Interaction Multivector Retrieval. - tusKANNy/awesome-multivector-retrieval

github.com
6
11
546
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoSep 23
Added dataset cards to all of these (plus the older MS MARCO v1 & LoTTE pooled) to make them truly plug-and-play. Turns out people can't read my mind to figure out how to use them 😮‍💨
Silvio Martinico@SilvioMartinicoSep 22
🧵 Just added 9 pre-encoded multivector BEIR datasets to Hugging Face (NFCorpus, SCIDOCS, FiQA across ColBERTv2, ModernColBERT, LateOn), with top-100 groundtruth and reference metrics. Perfect for benchmarking your methods without encoding overhead. #LateInteraction #NeuralIR
1
7
504
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoSep 22
🧵 Just added 9 pre-encoded multivector BEIR datasets to Hugging Face (NFCorpus, SCIDOCS, FiQA across ColBERTv2, ModernColBERT, LateOn), with top-100 groundtruth and reference metrics. Perfect for benchmarking your methods without encoding overhead. #LateInteraction #NeuralIR
1
2
1
19
1,187
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoSep 15
RaBitQ is now in Vectorium! Up to 30x compression, no training, 1/2/4/8 bits per component, 1.4-2.2x faster than the original implementation. Against PQ, on neural embeddings it wins on every axis, and the gap grows with dimensionality. #VectorSearch #Quantization #NeuralIR
2
6
17
781
Franco Maria Nardini retweeted
Connor Shorten
@CShorten30Sep 4
🔥 Information Retrieval in 2030 🔥 @lateinteraction at VLDB Extremely excited about OBLIQ and the call for more ambitious search benchmarks! 🚀 And who isn’t excited about the progress in late interaction engines? ⚡️
2
13
1
59
6,500
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoAug 10
It would be cool if someone had a retrieval library with SOTA performance on both dense and sparse vectors... oh wait, that's exactly kANNolo! 😁👇 arxiv.org/abs/2608.02583v1

UEmbed: Unified Sparse and Dense Multimodal Embeddings

Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical...

arxiv.org
1
4
13
716
The final program of ReNeuIR 2026 is available at reneuir.org/program.html. An excellent lineup of speakers and plenty of opportunities for engaging discussions. See you on Friday! #SIGIR2026

Tentative Workshop Program

Workshop on Reaching Efficiency in Neural Information Retrieval

reneuir.org
1
2
4
1,508
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoJul 16
I will present TACHIOM at #SIGIR2026 next week in Melbourne! 🦘 Exactly one year after I first discovered multivector retrieval at SIGIR 2025, I'm bringing a massive leap in efficiency and scalability to the very same conference. 🧵👇 #LateInteraction #NeuralIR #Efficiency
1
1
15
598
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoJul 15
Late interaction is absolutely exploding this year 🚀 We had a packed LIR workshop at ECIR 2026, and next week at #SIGIR2026 there is an entire oral session dedicated to multivector retrieval (Tuesday afternoon, don't miss it!) 🧵 #LateInteraction #InformationRetrieval
1
2
16
873
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoJul 13
People often ask me if they can feed their documents directly into kANNolo. Starting today, you can. kANNolo is now officially integrated into PyTerrier! 🚀🧵 #InformationRetrieval #NeuralIR #VectorSearch
1
4
1
19
1,260
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoJun 30
Why build another vector search engine? Because we wanted extreme modularity with no performance compromise. kANNolo beats FAISS & hnswlib on dense retrieval, and significantly surpasses NeurIPS '23 BigANN winners on sparse🚀 🧵👇 #VectorSearch #RustLang #InformationRetrieval
3
8
1
45
4,002
TACHIOM is now in PyLate: lightning-fast multi-vector indexing and search directly on CPU! arxiv.org/abs/2604.28142 Joint work with @SilvioMartinico, @cosimorulli1, @rventurini_. Thanks, @antoine_chaffin, and the PyLate team, for the support with the integration!

Efficient Multivector Retrieval with Token-Aware Clustering and...

Multivector retrieval models achieve state-of-the-art effectiveness through fine-grained token-level representations, but their deployment incurs substantial computational and memory costs....

arxiv.org
Antoine Chaffin
@antoine_chaffinJun 11
Whether you are GPU poor or GPU rich, today's release of PyLate has something for you! GPU maxxers: MaxSim kernels greatly speed up training while lowering the memory requirements CPU enjoyers: TACHIOM enables lightning fast multi-vector indexing and search directly on CPU
2
19
650
Franco Maria Nardini retweeted
Antoine Chaffin
@antoine_chaffinJun 11
Whether you are GPU poor or GPU rich, today's release of PyLate has something for you! GPU maxxers: MaxSim kernels greatly speed up training while lowering the memory requirements CPU enjoyers: TACHIOM enables lightning fast multi-vector indexing and search directly on CPU
3
29
6
126
12,879
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoJun 2
The late-interaction multivector retrieval ecosystem is exploding right now. To help separate the signal from the noise, we put together an "Awesome Multivector Retrieval" list organizing the top models, engines, libraries, and datasets all in one place 📚 🧵👇
5
27
1
119
7,470
Franco Maria Nardini retweeted
Silvio Martinico@SilvioMartinicoMay 31
Quick update: TACHIOM 0.3.0 is out with mean-centering to help alleviate the anisotropy problem. Also noticed that newer models usually need lower micro/small token thresholds than the defaults calibrated on ColBERTv2.0. More to come soon! ⚔️
1
4
1
21
2,314
Franco Maria Nardini retweeted
Antoine Chaffin
@antoine_chaffinMay 27
TACHIOM is very cool and is getting merged in PyLate very soon thanks to the amazing work of @SilvioMartinico He and the other people of the TusKANNy are shipping insane stuff (in sparse as well!) and are very nice 🫶
Silvio Martinico@SilvioMartinicoMay 27
Late-interaction retrieval is incredibly powerful, but scaling it is computationally challenging. k-means is a huge bottleneck. Our new architecture, TACHIOM, is fully open-source and tackles this problem: up to 247x faster clustering and 9.8x faster retrieval. ⚡ 🧵👇
1
2
38
2,491
Franco Maria Nardini retweeted
Omar Khattab
@lateinteractionMay 28
this is such an impressive result: search over ~600,000,000 colbert vectors in 10 milliseconds, with a *single* CPU core. and since this algorithm has sub-linear latency, there’s no excuse for anyone up to tens of billions of tokens
Silvio Martinico@SilvioMartinicoMay 27
Replying to @SilvioMartinico
For the final refine phase, we implemented a cache-optimized Product Quantization (PQ) layout specifically tailored for late interaction. Evaluated on ColBERTv2.0 embeddings, it results in 10 ms single-CPU retrieval on large-scale datasets (MS MARCO-v1, LoTTE Pooled).
9
33
7
323
48,201
Load more