MICES 2026

Opening and welcome by the MICES organizers René Kriegler and Sebastian Russ.

Precision vs. Recall: When Good Enough Beats Perfect

Arne Vogt & Kolja Hopfmann (OTTO)

Precision is a fundamental pillar of information retrieval. Yet we struggled to prove its value. Over the past years, we transformed our search from purely lexical to fully hybrid, where each query can now include results found through semantic retrieval. The greater fuzziness of semantic search proved very effective in closing recall gaps, with clear positive effects on conversion rate demonstrated in several AB tests. However, this fuzziness came at a cost: user feedback revealed a significant drop in result relevance, signaling that our precision had decreased — and this uncovered a deeper problem: How does higher precision affect user behavior, and which metric actually demonstrates its added value?

We experimented with dynamic similarity thresholds, query intent based precision optimization and post-filtering on certain attributes, testing against both economic and effort metrics. None of them showed a clearly positive effect in AB tests. Is relevance simply non-binary, as suggested by Amazon's ESCI logic? Are the benefits of precision optimization only visible in long-term effects that AB tests cannot capture? Or are users just more tolerant of imprecise results than we assume? In this talk, we present our journey and challenge the community to rethink how precision should be measured in modern e-commerce search.

Arne is Lead Product Manager for Search & Discovery at OTTO, one of Europe's largest e-commerce platforms. With over fifteen years of experience in e-commerce search, he bridges the gap between user experience, business impact and technology. Recently he has also been heavily involved in building OTTO's conversational search assistant, exploring how dialogue-based search changes the way users find products.

Kolja is a software engineer based in Hamburg, specializing in Natural Language Processing and search technologies. With a strong background in Python and Java, they focus on building intelligent, user-centric systems that enhance digital experiences.

Search, Chat, Ockham: When Two Interfaces Are One Too Many

Andrea Polonioli (Coveo)

Over the past twelve months, ecommerce chatbots have gone mainstream — from Macy's to Argos, from Vans to Belk and now Frasers — positioned as the natural, expressive alternative to the rigidity of keyword search. Meanwhile, search itself has quietly transformed: queries that once broke traditional engines — multi-constraint inputs, situational requests, price-qualified natural language — can now be handled with reasonable accuracy. Both channels are converging on the same capability, and yet most large retailers still deploy them as separate systems with separate logic, separate pipelines, and no shared governance over conflicting results for the same shopper intent.

This talk applies Ockham's razor to experience design: do not multiply interfaces without necessity. Maintaining parallel search and chat surfaces is justified only when each offers a distinct, predictable advantage, or when both are coordinated enough to behave as one — in practice, neither condition holds, and the result is duplicated logic, fragmented metrics, and a discovery experience that is ungovernable almost by design. There are contexts where a bounded conversational agent earns its place precisely because its scope and purpose are clearly defined; that clarity is exactly what the generic sidecar widget lacks. The conclusion: product discovery must be treated as a single intelligence, not two systems that happen to share a catalog — and a shopper.

Andrea is an expert in ecommerce search and product discovery. He works at Coveo and previously at Tooso, a Gartner Cool Vendor in Digital Commerce. He is a widely cited author with hundreds of citations and 15+ peer-reviewed publications.

The Journey to Semantic Search in Omnichannel Retail at dm-drogerie markt

Denise Schäfer (dm-drogerie markt) & Mike Dirnberger (diva-e)

Semantic search is transforming product discovery in e-commerce by moving beyond keywords toward a deeper understanding of user intent. We share how dm-drogerie markt has been building a semantic search system that improves relevance, reduces zero-result queries, and scales to millions of daily searches — outlining the gap between keyword search and semantic retrieval, key design choices for production-ready vector-based retrieval, and how offline evaluation, LLM-supported relevance checks, and A/B testing guide the process.

Practical examples from real usage scenarios illustrate our iterative approach: introducing guardrails such as relevance thresholds and category restrictions, incorporating business signals, and adding brand-aware and attribute-based filtering for trustworthy results. We also highlight concrete use cases including zero-result handling and multi-stage intent detection for chatbot and MCP integrations, offering actionable insights for building reliable semantic search systems in large-scale retail environments.

Denise currently works in the team responsible for search in the dm-drogerie markt online shop and app. She has more than 25 years of experience in software development and project management, and is passionate about all things technological, architectural, and innovative.

Mike is a backend software engineer and search consultant at diva-e, helping clients design, implement, and optimize search solutions – from open-source to proprietary – using a data-driven approach. He has been part of a dedicated e-commerce search team at dm-drogerie markt for the past five years.

Inside Zalando Search: Architecture Behind Product Discovery at Scale

Ivan Potapov (Zalando)

A look inside the Search & Browse architecture powering product discovery for millions of customers across Zalando's European markets. We'll walk through the system end-to-end: Elasticsearch clusters with lexical and vector retrieval, the NER-powered query builder that interprets user intent, the Catalog API orchestrating parallel requests, and the ranking stack — where the Algorithm Gateway applies ML relevance models and re-ranking to turn raw search results into a curated shopping experience. We'll also touch on how promotions blending and multi-market isolation fit into the picture.

With a product perspective on what these architecture and ranking choices mean for search quality, partner visibility, and the customer experience. Follow up talk on blog post: engineering.zalando.com

Ivan is a Research Engineer at Zalando on the Search & Browse team, powering catalog search and product discovery for millions of customers across Europe. He has a deep interest in Code Agents architecture, is an active open-source contributor, and writes at blog.ivan.digital.

Fine-Tuning Sparse Neural Retrievers for E-Commerce Is Not That Scary (And Often Worth It)

Evgeniya Sukhodolskaya (Qdrant)

Sparse neural retrieval is the middle ground between BM25+LTR, which you've already squeezed, and dense retrieval you'd rather not put in front of users. Inverted index, exact term matching, often synonymical term expansion... SPLADE is a known example. It pays off once fine-tuned, and that's where many get scared.

This talk is to face your fears: the minimal pipeline (SPLADE training, hard negative mining against a vector index, evaluation), the specialization-vs-generalization tradeoff, and a framework that runs the pipeline end-to-end. Let's see when sparse neural is worth trying, when fine-tuning helps, and a practical way to run it all end-to-end.

Evgeniya is a Senior Developer Advocate at Qdrant with 8 years of IT experience across software engineering, machine learning, and technical management, and 4 years in Developer Relations. Holds a Master's in Machine Learning, Data Analytics, and Data Engineering. Passionate about NLP, data-centric AI, and the role of vector search in advancing AI technologies.

How much searchandizing is too much searchandizing?

Charlie Hull (The Search Juggler)

You've installed a new search engine with powerful matching and ranking features — and now your executives, commercial team and users are complaining that certain products don't appear, suppliers are demanding top placement, and common misspellings lead to zero results. Searchandizing to the rescue! You start adding rules, redirects, query rewrites, synonyms and hacks…. Months or years later, the original search logic is buried beneath so many special cases that you're not sure anymore why anything appears where it does. Maybe you've gone too far.

This talk covers when searchandizing should — and should not — be used, which techniques to apply, and how to make results explainable both internally and externally. It will describe processes and good practices to manage searchandizing proactively rather than reactively, how to prioritise, how to push back when nothing should change, and why ongoing maintenance and the ability to review and remove old rules matters.

Charlie is a leading figure in the search industry, known for an honest, neutral and pragmatic viewpoint. He has held multiple roles including senior consultant, strategic advisor, project manager, sales & marketing director, conference organiser & speaker, trainer, writer & mentor. He is deeply connected with the business & technology of website and enterprise search engines with particular experience of small, high-value consulting companies and open source stacks.

Semantic + Lexical Retrieval for Dynamic Sponsored Search

Kunal Sonalkar (Nordstrom)

Hybrid retrieval combining SigLIP (vision models) and BM25 (lexical model) is especially useful in sponsored search, where ad inventory changes frequently as budgets, bids, and eligibility shift. Because the set of available ads is dynamic, ranking cannot rely only on historical popularity or static matching — it must depend on content-based relevance. SigLIP captures semantic and visual similarity while BM25 preserves exact keyword matching, creating a more robust retrieval stack for sponsored results. The session will further explore how to leverage LLMs to generate better product descriptions for improved BM25 scores, and how a query intent model can sharpen sponsored search relevance.

Kunal is a data scientist at Nordstrom, working at the intersection of search relevance, sponsored search, and recommender systems.

Hybrid search at idealo: from fine-tuning to production

Gennady Shabanov (idealo)

We share our journey of bringing hybrid search to idealo — combining vector and keyword retrieval in production. We cover how we fine-tuned our embedding model, the challenges we faced, and key observations along the way, including how we leveraged LLMs in the fine-tuning process, and how we built an in-house vector search index that searches across 500 million vectors with an average latency around 65ms.

We then compare two approaches to fusing keyword and vector results: Reciprocal Rank Fusion (RRF) and Learning-to-Rank (LTR) as a reranker over the combined candidate set. We discuss how we approach offline evaluation, its challenges, and how we use LLMs to support it, share the trade-offs of each fusion approach, what the A/B tests showed, and close with the measurable impact hybrid search had on our core KPIs: CTR, clickouts, and exit rate.

Gennady has been a Machine Learning Engineer with idealo GmbH's search team for the past five years. In this role, he is responsible for developing and implementing Machine Learning solutions and the supporting infrastructure within idealo's search platform. Prior to joining idealo, he worked at Ladenzeile.de, where he contributed to a large-scale classification system for e-commerce products.

Why your B2B Search Engine doesn't understand your users?

Maëlly Dubois (Adelean)

E-commerce search engines are often optimized for simple, "product-centric" queries. But users search differently, especially in B2B: they describe their need, their usage, and their constraints with highly specialized (and sometimes long) business queries. The result: zero results, irrelevant products, degraded relevance… even though the right products exist in the catalog.

In this talk, we start with a real-life large-scale B2B e-commerce search case and show that the problem is not "ranking," but a combination of overly strict retrieval, poor query understanding, a single search strategy for multiple intentions, and product data poorly aligned with real-world usage.

Through a decision-making diagnostic tree, we will see how to precisely identify where the search fails (strict AND, stopwords, field weighting, routing by intention, data normalization), and then how to design targeted experiments to improve relevance without "starting from scratch."

Maëlly is a Product & Operations Manager at Adelean, specializing in search, data, and complex systems, with a particular focus on e-commerce search engines. She works on topics that combine architecture, relevance, product data, and technical decision-making. This dual product and consulting background allows her to bring a strong field-oriented perspective, grounded in real-world data and actual usage patterns.