High Performance Computing Laboratory · ISTI-CNR · Pisa
We design algorithms and systems that make information access and AI fast, practical and scalable under real-world constraints: massive data, tight latency, limited energy and hardware budgets, distributed infrastructures and privacy. Our work spans efficient retrieval over learned sparse and dense representations, learning to rank and conversational search; semantic enrichment of mobility and social data; resource-aware, decentralized AI, and orchestration across the cloud–edge continuum. We pair algorithmic design with careful engineering and reproducible evaluation.
Research
LLMs and agentic applications that stay accurate and reliable under limits on latency, memory, energy, bandwidth and privacy. Compression, quantization, distillation, pruning, caching and adaptive inference; parameter-efficient fine-tuning; federated and decentralized AI for services that run close to users and data, across cloud, edge and device.
Accurate, low-latency search over massive, heterogeneous collections. Learned dense and sparse representations and their indexing, compression and querying; vector search and approximate nearest neighbours; learning to rank, re-ranking, hybrid retrieval and retrieval-augmented generation that balance effectiveness and efficiency.
Platforms that bring computation and data closer to users across cloud, edge, device and federated environments. Decentralized, AI-enhanced orchestration for service placement, scheduling, resource allocation and migration that adapts to workload, resources and mobility, cutting latency, bandwidth, energy and cost.
AI-driven analysis of mobility, social, textual and spatio-temporal data. Semantic trajectories, large-scale mobility analytics, context-aware services and LLM-based understanding of text and social media, for smart mobility, transportation, tourism and urban computing.
People
Many of our alumni now hold positions at universities (Sapienza University of Rome, Ca' Foscari University of Venice, UCL) and companies such as Meta, Twitter, Bloomberg, Ferrari, Yahoo and DuckDuckGo.
Results & impact
Award-winning work, patented technology and our most-cited papers of recent years.
D. Dato, C. Lucchese, F. M. Nardini, S. Orlando, R. Perego, N. Tonellotto, R. Venturini · Assignee: Istella S.p.A.
US 11,106,685 B2 · WO 2016/203501
O. Frieder, I. Mele, R. Perego, N. Tonellotto · Assignee: Georgetown University
US 10,503,792 B1 · US 11,151,167 B2 · US 11,693,885 B2 · WO 2020/231382
O. Frieder, I. Mele, C. I. Muntean, F. M. Nardini, R. Perego, N. Tonellotto · Assignee: Georgetown University
US 12,067,021 B2 · continuation pending (US 2024/0370446 A1)
Citation counts from OpenAlex, updated with each build.
Software & data
Our research ships as code and data others can reuse. Most tools are open source; our efficient-retrieval libraries are collected under TusKANNy ↗.
tusKANNyRust · learned sparse retrieval
Fast retrieval over learned sparse embeddings. Inverted lists are split into geometrically cohesive blocks with summary sketches, so most blocks can be skipped at query time. SIGIR 2024 Best Paper Runner-up; 1st and 2nd place at the SISAP 2026 Indexing Challenge with kANNolo.
tusKANNyRust · ANN search
A research library for approximate nearest-neighbour search over dense and sparse embeddings, built for easy prototyping without giving up speed. Implements HNSW graph indexes and product quantization.
C++ · learning to rank
An efficient learning-to-rank toolkit with C++ implementations of GBRT, LambdaMART, Oblivious GBRT/LambdaMART, Coordinate Ascent and RankBoost.
Python · ranking evaluation
Analysis and evaluation of learning-to-rank models built on ensembles of regression trees: interoperable tools to compare and dissect models from different LtR libraries.
Web app · tourism
Plans time-budgeted sightseeing tours from points of interest and tourist movement patterns mined from Wikipedia and Flickr. ECIR 2014 Best Demo Award.
Web search · query suggestion
Query suggestion that copes with data sparsity, so it works well for rare queries in the long tail of the query distribution.
Awards
When Reducing Representations Improves Performance
SISAP 2026 Indexing Challenge, Task 3 (sparse)
Investigating the Scalability of Approximate Sparse Retrieval Algorithms to Massive Datasets
Efficient Inverted Indexes for Approximate Retrieval over Learned Sparse Representations
A Geometric Framework for Query Performance Prediction in Conversational Search
Ensemble Model Compression for Fast and Energy-Efficient Ranking on FPGAs
Xilinx University Donation Program
NVIDIA Academic GPU Grant Program
SEL: A Unified Algorithm for Entity Linking and Saliency Detection
QuickScorer: a Fast Algorithm to Rank Documents with Additive Ensembles of Regression Trees
Partitioned Elias-Fano Indexes
TripBuilder: A Tool for Recommending Sightseeing Tours
Yahoo Faculty Research and Engagement Program (two awards)
Baquara: A Holistic Ontological Framework for Movement Analysis Using Linked Data
Discovering Tasks from Search Engine Query Logs
Monte Carlo Simulation for SPECT with XtreemOS
Identifying Task-based Sessions in Search Engine Query Logs
Efficient Dynamic Pruning with Proximity Support
Mining Top-K Patterns from Binary Datasets in Presence of Noise
ERCIM Cor Baayen Award
Sorting Out the Document Identifier Assignment Problem
An Online Recommender System for Large Web Sites
Contact
We welcome Ph.D. candidates, visiting researchers and industrial partners interested in efficient search, AI and data systems. Write to the lab head to get in touch.
From Pisa Centrale, take bus LAM Verde towards Via di Pratale and get off at the first stop in Via Volpi. The CNR campus is a short walk away.
