Beyond Text: Imperfect Labels and Intent in Visual Retrieval

Visual retrieval is moving beyond text. Join us for two technical talks: teaching models to search documents by treating pages as images, including the problem of imperfect labels in training, and using intent detection to build image discovery in a fashion AI assistant. Pizza and drinks on us — stick around after for networking.
✍️ RSVP soon — capacity is limited!

📅 Date and Time
Wednesday, November 11, 2026 from 18:00-20:45

📍 Location: Hedwig-Wachenheim-Straße 7, 10243 Berlin
BHW Auditorium (1st floor), Zalando SE
📝 Agenda
18:00 – Doors open, arrivals and check-in
18:15 – Welcoming and open session
18:30 – Teaching Models to Search Documents: What If the “Wrong” Answer Is Right? by Saba Sturua, Senior AI Research Engineer @ Elastic
Finding information in PDFs means navigating tables, charts, and page layouts as well as text. Visual document retrieval tackles this by treating pages as images. But there’s a subtle problem: the training data may recognize only one correct answer, even when several pages answer the same question.
In this talk, I’ll introduce visual document retrieval and explore how larger models can help train smaller search models. Drawing on our recent research, we’ll compare teaching strategies, from highlighting relevant regions to judging which retrieved pages answer a question. In our experiments, some of the most useful supervision came from helping the retriever avoid learning that a valid answer was wrong. This work is part of the research behind our next generation of Jina embedding models.
Through concrete examples and experiments, we’ll explore why seemingly good training examples can backfire, and what this tells us about building search systems with imperfect labels.
19:00 – Q&A and quick break
19:15 – From Conversation to Visual Search: Building Image Discovery for Zalando AI Assistant by Inanda Menezes, Principal Software Engineer @ Zalando
How can a fashion AI assistant understand a customer's intent from visual input? In this talk, we’ll explore how we built image discovery for Zalando’s Assistant, considering the intent detection to support different use cases, from finding similar products to getting fashion advice. We’ll dive into how we connect these intents to the right processing and retrieval strategies, and share the key challenges.
19:45 – Q&A, pizza, drinks & networking
\
Invite your friends and join us for an evening of learning and networking with fellow engineers and speakers!

Want to speak at our next meetup? Submit your talk on this link and we will get in touch with you as soon as possible.