Join us in Berlin for Haystack Europe 2026 – an immersive experience featuring workshops and two days of conference talks tailored for search and AI professionals.
Search is no longer just a product feature — it’s becoming a foundational platform that powers AI assistants, copilots, personalization engines, and internal knowledge systems. Shifting from a solid onsite search experience to a true search platform raises new questions: how should teams collaborate, what data needs to be exposed and in what form, and how do you design for both human users and AI agents at once?
Large Language Models (LLMs) transform how we build and evaluate search systems, it’s crucial to understand how to use them effectively as “judges.” This condensed, hands-on training introduces the principles and practical techniques for implementing “LLM as a Judge” to evaluate search result quality.
This talk presents a practical evaluation framework for an e-commerce conversational search assistant. In addition to a lightweight LLM-as-a-judge, we track additional metrics such as search term accuracy, filter precision and no-results rate. Learn how these signals surfaced concrete failure modes, guided our iterations and improved relevance.
We trained an open image-text embedding model in hyperbolic space, so visual search can use hierarchy as well as similarity. The talk shows how radius and aperture create hierarchy inside the embedding model, how that helps retrieve specific items from compositional queries, and where the approach fails.
Netflix search has evolved from simple Trie-based lookups to complex, intent-driven conversational queries. This talk explores our journey to integrate deep Query Understanding and semantic retrieval at scale.
TNG's R1T Chimera models reached over 10 billion daily tokens on OpenRouter. How can a small consultancy produce its own LLMs? In our case: by Frankensteining them. This talk motivates: you can adapt LLMs to your own needs without being a big research lab. Including theory, technicalities, and lessons learned.
Think larger context window solves enterprise AI? Think again. Learn how recursive agentic retrieval progressively constructs the right context for every reasoning step, enabling AI agents to solve complex enterprise tasks beyond traditional RAG.
Balancing business rules, vector search, and AI evaluation is a shared challenge. Modern search systems must now serve both human and agentic users. In this talk we share grounded insights on search replatforming, API-driven tools, deterministic query governance, and the growing complexities of evaluation in the age of generative AI.
Search is essential to how shoppers discover products, and we asked whether machine-learned re-ranking could improve search effectiveness across hundreds of commerce clients. This is that journey: from a baseline and a null A/B test, through modelling, architecture, and evaluation, to a client revenue win, and toward mass generalisation.
Modern IR techniques like Late Interaction (multivector representations) and wormhole vectors (hopping between sparse & dense vector spaces) newly expand our toolbox for building state of the art Query Understanding. We'll show you how, including ranking benchmarks that blow away today's popular hybrid search (BM25 + vector similarity) techniques.
Discover how Leroy Merlin bridges product catalogs with Knowledge Graphs using Sparse Encoders. We’ll share how we injected KG concepts as model tokens, tuned custom layers for 99%+ sparsity, and achieved explainable, low-latency (60ms) semantic search indexing ready for Elasticsearch.
At OTTO, we want conversational AI to become the primary interface for product discovery.
We built a weekly loop that mines production sessions, grounds an LLM judge in product discovery signals, clusters failures, and ships fixes to search tools and orchestration. In this session, you'll get the building blocks, our key findings and outcomes.
"Just let the agent search live" has a cost curve, so does a compressed index. The curves cross but most teams pick a side before doing the maths. We'll use open benchmarks on quantisation and refinement to prove what compression really costs you in recall, in answer quality, and in dollars against live search.
Drawing on behavioral UX, cognitive offloading and spatial thinking, Öykü will share practical design principles for making AI-powered search easier to understand, inspect and challenge: How can interfaces surface sources, uncertainty and alternative perspectives while helping people remain active participants in the decision-making process.
We all know what happens when AI learns from unrepresentative data. We recognize the problem, measure the bias, try to mitigate it — and eventually put a human in the loop. Problem solved, isn't it?
This talk turns the usual human-in-the-loop question around and asks whether (and how) we should also debug the human factor.
We introduce per-document fine-tuning, a method that updates individual embeddings with new information while keeping the embedding model frozen. For example, applied to product search, it enriches product image embeddings with abstract attributes such as material and use case, improving retrieval without costly and risky model fine-tuning.
An agent that forgets you between sessions breaks a core UX promise. For our Agent Studio, we built memory so that it works across classic search, vectors, or both, based on what you already have. I'll share our journey: what worked, what surprised us, and why 65% is sometimes the best score you can get.