Talks

Title:
GaussVector: A Multi-modal Vector Database for the Agentic AI Era

Abstract:
As large language models and agentic AI systems become central to enterprise operations, production deployments demand vector database infrastructure that is simultaneously fast, accurate, scalable, and cost-efficient — with seamless integration across diverse application ecosystems. GaussVector is Huawei’s production-grade distributed vector database, purpose-built for multi-modal, enterprise-scale AI workloads. Its storage-compute separation architecture achieves sub-second retrieval over 100B-scale datasets with 10× storage capacity improvement and 100×+ faster node scaling over traditional deployments. Leveraging Huawei’s Ascend NPU co-design stack, GaussVector further delivers 10× index build speedup and 4× query throughput improvement over CPU baselines. GaussVector is a unified multi-modal storage engine consolidates relational, vector, and graph workloads in a single system, and an open ecosystem design enables seamless integration across the full AI stack.

Bio:
Huaxin Zhang is a Senior Principal Software Engineer with two decades of industry experience in database systems and data infrastructure. He holds a Ph.D. in Computer Science from the University of Waterloo, spent 10 years at IBM working on DB2, followed by 10 years at Huawei advancing GaussDB. Huaxin holds 10 U.S. patents, and is currently focusing on vector database architecture and storage engine design.

Title:
Semantic Query Processing over Relations

Abstract:
Language models are making it possible to ask richer questions over relational data, but doing so efficiently remains difficult. Join-heavy queries, often over networked data, can produce large intermediate results that must be serialized into prompts and then fed into language models. This talk presents FFX (Fast Factorized eXecution), a query engine that combines factorized and vectorized execution to address this bottleneck.

The talk focuses on how FFX changes semantic query processing by keeping join intermediates compact, enabling semantic operators to serialize factorized intermediates and predict over their implied Cartesian products. Operators then produce predictions as flat output tuples and bypass having to first flatten the input relation. Empirically, and somewhat surprisingly, our evaluation shows that even non-reasoning models can often perform this Cartesian expansion accurately while still carrying out the semantic task. In our evaluation, FFX achieves an order-of-magnitude reduction in input tokens while maintaining the same accuracy and degrades more gracefully as context size increases.

Bio:
Amine Mhedhbi is an assistant professor at École Polytechnique de Montréal. His interests are in building and analyzing analytical and AI-driven data system architectures. His work includes tackling performance considerations, debuggability, interface design, and data applications. Amine received his Ph.D. in 2023 from the University of Waterloo. His research has been awarded a VLDB best paper award, a Microsoft Ph.D. fellowship award, and the University of Waterloo’s Computer Science distinguished dissertation award.

Title:
Data in the Age of AI

Abstract:
The traditional data lifecycle has involved gathering functional requirements, gathering approvals, building a data stack, and maintaining the stack as application needs change. To what extent does this all change in the age of AI? I will point to some challenges in using AI to automate parts of this stack.

Bio:
Kavitha Srinivas is a senior research scientist at IBM. Her research interests have spanned semantic web, knowledge graphs, AI planning, and more recently the application of AI to data management problems. In particular, she is interested is in using agents to semi-automate data management tasks.