The future of RecSys: vector databases, enhanced calibration, explorative search

## RecSys IL Meetup #5

The future of RecSys: vector databases, enhanced calibration, explorative search.

Talks are in English.
We are grateful to eBay for their community sponsorship and to Taboola for hosting this event.
Zoom link for those joining online:

Agenda:
18:00 - 18:30 Registration, Mingling, Refreshments & Beer
18:30 - 19:00 "Engineers approach to building real-time RecSys" (Daniel Svonava, Superlinked)
19:00 - 19:30 "Enhancing model calibration in recommender systems" (Hai Sitton, Taboola)
19:30-20:00 "NERetrieve, Exploration and the Future of Search" (Uri Katz, Bar Ilan University)

Abstracts:
1. Engineers approach to building real-time RecSys" (Daniel Svonava)
There are now 37 vector databases on the market - we counted.
But how do people actually use these things to build useful systems?
Let’s look at a practical RecSys implementation we deployed for an e-commerce company with 2 million users with Superlinked and Redis.
Short bio:
Daniel is an entrepreneurial technologist with a 20-year career. After 6-year tenure as a tech lead for ML infrastructure at YouTube Ads, Daniel co-founded Superlinked.com - an ML infrastructure startup that makes it easier to build information-retrieval-heavy systems - from Recommender Engines to Enterprise-focused LLM apps

2. Enhancing model calibration in recommender systems" (Hai Sitton)
Taboola’s recommender system is powered by a clickthrough rate (CTR) and conversion rate (CVR) estimations. Ensuring model calibration is vital in developing recommender systems for various real-world applications, yet it often receives insufficient attention. In this talk, we introduce how to measure model calibration and discuss our efforts in Taboola to make the models more calibrated.
Short Bio:
Hai Sitton is an algorithm engineer in Taboola

3. NERetrieve, Exploration and the Future of Search (Uri Katz)
The ability to retrieve documents based on specific entity attributes (e.g., types of viruses affecting certain brain regions, or details of tech company acquisitions) remains a significant challenge for search engines. We introduce NERetrieve, a new paradigm combining Named Entity Recognition (NER) and Information Retrieval (IR). To support this work, we present a massive dataset and demonstrate how even state-of-the-art semantic search methods struggle with the precision required for this task. NERetrieve opens a new frontier in how we interact with information, enabling highly targeted and nuanced knowledge discovery.
Short Bio:
Uri Katz is a PhD student at Bar-Ilan University, exploring information retrieval and knowledge discovery in Yoav Goldberg's lab. His research focuses on improving how we find and use information.