Bavaria, Advancements in SEarch Development (BASED) Meetup
How to Find the Venue
The IntraFind office is located at Landsberger Straße 368 in the Laim district (west of Munich). The main entrance is on the right-hand side at the rear of the office building.
About us
Search is evolving fast. From new algorithms and tools to AI-driven solutions, the field is constantly shifting.
BASED Meetup is where professionals & enthusiasts discuss the latest trends, breakthroughs, and challenges in modern search.
We dive into topics like:
- Agentic Retrieval-Augmented Generation (RAG);
- Innovative approaches to sparse, dense, and hybrid retrieval;
- Advances in indexing algorithms and vector quantization;
- Modern metrics for search quality;
- Optimization techniques for production-scale systems;
- Cross-lingual and multimodal search challenges;
- The intersection of search and generative AI;
… and much more.
No sales pitches, no gatekeeping — just an open space to share ideas, learn from each other, and explore the technologies shaping the future of search.
Whether you're building production systems, researching search algorithms, or just curious about the field, you're welcome here.
Agenda
18:00 — Doors Open;
18:15 - 18:30 — Welcome Note (Daniel Wrigley, Evgeniya Sukhodolskaya, IntraFind);
18:30 - 19:00 — "Maybe Dense Vector Search isn't as good as everybody thinks" (Christoph Goller, IntraFind)
19:00 - 19:10 — Break
19:10 - 19:40 — "Lessons Learned from Building an App for Navigating Bureaucracy" (Phan Van Nguyen Duc, Saclora UG)
19:40 - 20:10 — "LLMs as inexpensive raters: Setting up offline search evaluation infrastructure on a budget" (Mikayla Webster, HSE Home Shopping Europe GmbH)
20:10 - 20:15 — Closing Note (Daniel Wrigley, Evgeniya Sukhodolskaya)
20:15 - 21:00 — Networking with drinks & pizza sponsored by IntraFind
Talk Descriptions:
Talk 1: Maybe Dense Vector Search isn't as good as everybody thinks
Speaker: Christoph Goller (IntraFind)
Abstract: Recently, several prominent studies have questioned the effectiveness of dense vector search. Research from Prof. Schütze’s group provides empirical evidence that dense retrieval models often rely on superficial cues—such as exact string matches, repetitive content, or information appearing early in documents—rather than genuine semantic understanding or answer relevance. Meanwhile, work from Johns Hopkins University and Google DeepMind highlights theoretical limitations of dense vector search. I would like to present these findings and initiate a discussion on the future of semantic search.
Talk 2: Lessons Learned from Building an App for Navigating Bureaucracy
Speaker: Phan Van Nguyen Duc (Saclora UG)
Abstract: Filio is an app developed to help immigrants in Germany with a “simple” goal in mind: to provide people with the information they need, when they need it.
The project began with experiments using chatbots, retrieval-augmented generation (RAG) systems, and knowledge graphs to automate information delivery. Over time, it evolved into a community-driven platform where users can share verified knowledge and experiences.
In this talk, I will share my learnings about the limitations of each approach and provide a glimpse into the challenges of German bureaucracy and providing the right information to a highly diverse user base.
Talk 3: LLMs as inexpensive raters: Setting up offline search evaluation infrastructure on a budget
Speaker: Mikayla Webster (HSE Home Shopping Europe GmbH)
Abstract: In order to improve search result quality, we need to know what “good” results look like in the first place. Standard evaluation structures that rely on user feedback or industry expert judgements can be slow, labor intensive, and expensive. As new features and documents are incorporated into our search engines, the cost of constantly re-assessing relevance can become a barrier to iterative improvement. If you are lacking in budget, time, or workforce, Large Language Models (LLMs) can be a powerful tool to simulate expert judgements cheaply, quickly, and with comparable quality. In this talk, I will walk you through an e-commerce use case for setting up automated offline evaluation infrastructure to measure recall and precision using LLM-generated judgement scores.
Venue
We're hosted this time by IntraFind, who also sponsors networking drinks and snacks.