OpenSearchCon Europe 2025

Search engines take in a query and return a set of results that best satisfy the need of the searcher. Words from a text box, user interface facets, and some user identity information have been the clues for good retrieval. Today, Large Language Models (LLMs) produce vectors with semantic information that augments or replaces lexical information for retrieving best results. OpenSearch's low-level APIs can connect with and use LLMs as part of an overall search solution. A full solution still requires some higher-level coding to take full advantage of these APIs. In this talk, we’ll cover some theory on similarity, and how and where vectors can help you retrieve better results. We’ll dive deep on opensearch-ml-quickstart (https://github.com/Jon-AtAWS/opensearch-ml-quickstart), an open-source example set for managing connections with OpenSearch, local models, and vector embeddings. We’ll walk through examples and code to help you get started quickly with OpenSearch’s vector capabilities: exact nearest-neighbor matching, dense vector embeddings, and sparse vectors. You’ll leave with a great tool in opensearch-ml-quickstart, along with the knowledge of how to apply it.

Forget the old search paradigms—AI is here to change everything! By blending sparse (traditional term-based) and dense (vectorized) search methods, OpenSearch creates a hybrid solution that optimizes search results, from precision to performance.

Learn how to implement hybrid search, combine neural embeddings from LLMs with traditional search techniques, and enhance scalability. This solution bridges the gap between legacy systems and AI-driven search, improving accuracy and relevance while scaling to enterprise-level needs. Transform how you search, analyze, and scale with OpenSearch.

There are lots of statistics we can calculate for search relevance: NDCG, precision, MAP, …. OpenSearch 3.0 calculates and displays them for you.

But metrics don’t tell the whole story. In particular, NDCG can’t give you the qualitative analysis that one algorithm is great for author names, and a different one is good for shapes and sizes. For that, you need eyeballing, that is, human intuition and judgment.

That’s why we’ve built tools to help you dig deeper into your search results. The Search Comparison tool in OpenSearch 2.11 is just the starting point – in 3.0, it lets you define search configurations and test sets of multiple queries and determine which queries’ results have changed the most… and then zoom in to the individual cases. You can integrate data on judgments and query frequency from User Behavior Insights for deeper analysis.

In this talk, we’ll discuss how traditional statistical metrics are complemented by eyeballing, how the new tools in OpenSearch 3.0 support disciplined eyeballing, and how you can use them to improve the quality of your search results.

Search engines guide users from their initial goals to the information that fulfills those goals. The retrieved items are deemed relevant when they align with the user's intent. Ensuring relevant results involves having the data in the catalog, analyzing it accurately, representing it through text and vectors, ranking it appropriately, and performing other actions.
In this session, discover how auction houses utilize vector search, powered by Amazon OpenSearch Service, to find similar items within their extensive database, enhancing the customer experience and preventing issues with counterfeit artworks, and see how it assists users in obtaining the most relevant results.
This session also uncovers some peculiarities in getting embeddings with AWS Bedrock and AWS SageMaker serverless inference for the abovementioned goals.

In this session, Uber, a premier member of the OpenSearch Foundation, will share how the unique demands of their low-latency, high-throughput search platform have driven impactful contributions to the OpenSearch project.

The discussion will explore Uber’s collaboration with the OpenSearch community to integrate critical learnings and innovations from their own search platform. Key highlights include the development of a modular, cloud-native architecture with clear separation between readers and writers, a high-performance messaging protocol for inter-service communication, and a pull-based indexing model.

These advancements, along with contributions from the broader OpenSearch community, are shaping OpenSearch into a cutting-edge, serverless open-source search engine tailored for enterprise needs. Join this session to discover how pioneering work from Uber and OpenSearch is paving the way for the next generation of high performance and scalable search technology in open source.

Dive deep into OpenSearch workload management and optimize your cluster's performance and resource allocation. This talk is suitable for both seasoned users and newcomers to OpenSearch.

We'll cover workload management fundamentals, exploring core concepts and principles specific to OpenSearch environments. You'll discover the benefits through real-world scenarios. We'll also guide you through implementation strategies, offering best practices and helping you avoid common pitfalls.

Gain insights into the various resource types supported by OpenSearch's workload management feature, enabling informed allocation decisions. We'll update you on current capabilities and provide a glimpse into the exciting future roadmap of this feature.

By attending, you'll gain actionable insights for immediate application, learn to address common OpenSearch challenges. This talk is ideal for OpenSearch administrators, DevOps professionals with large-scale search and analytics, IT decision-makers, and open-source enthusiasts.

Join us to master WLM in OpenSearch. Enhance your skills, optimize your clusters, and become part of the vibrant OpenSearch community driving innovation in search technology.

With NiFi version 2.x, the ability to interface with OpenSearch using pre-built Python components (processors) was introduced.

These processors address various needs related to AI solutions, specifically vector embedding using tools like ChatGPT or HuggingFace. The selection and implementation of these components were partly driven by the increasing interest in these topics in recent times.

In the presentation, we will showcase an ETL flow for populating an OpenSearch index with vector fields. The flow is built using NiFi for both data acquisition and vector embedding generation, as well as querying via the dedicated OpenSearch components.

A growing trend in recent times is the demand for AI systems that use local engines, independent of external access.

As an example of possible application, we will present a solution based on customizing PutOpenSearchVector QueryOpenSearchVector processors to perform vector embedding and search using local LLMs based on Ollama.

In this session, we'll delve into the critical optimizations and fine-tuning strategies that propelled HDI Germany P&C and Life Insurance, a Talanx AG subsidiary to develop a high-performance, productionized Retrieval-Augmented Generation (RAG) based GenAI solution. Leveraging OpenSearch with hybrid search, our system now empowers 300+ employees. Our journey from concept to large-scale deployment uncovered numerous minor yet powerful tweaks that significantly enhanced both, speed and accuracy of our Opensearch indexing and retrieval mechanisms. We’ll dissect the key parameters crucial for optimal RAG performance, presenting benchmarks and metrics to quantify these improvements achieved through the optimization process.

Attendees will gain practical insights into maximizing OpenSearch's potential for RAG applications, learning how to leverage these often-overlooked adjustments to achieve substantial performance gains.
Join us for an in-depth exploration of the nuances that can make a significant difference in your OpenSearch-powered RAG solutions, and leave with actionable tips to implement in your own projects.

In the era of vector search and semantic similarity, returning highly relevant results is only half the battle. When search results are too similar, users must wade through redundant information to find diverse perspectives. This talk introduces Maximum Marginal Relevance (MMR) implementation in OpenSearch, a powerful technique that optimally balances result relevance with diversity.

Key takeaways will include:
- Implementing MMR reranking with OpenSearch's vector search
- Optimizing performance for large-scale deployments
- Measuring and tuning diversity metrics
- Real-world applications and success patterns

With the rapid rise of Large Vision Language Models, radically new approaches to multimodal information retrieval have been proposed. Methods like ColPali and Document Screenshot Embeddings bypass the classical document preprocessing pipeline by representing document pages only in terms of vision based image patches. This allows unprecedented accuracy for Retrieval Augmented Generation (RAG) in answering questions about figures, tables, charts and visually complex documents like Powerpoint slides, without the hassle of complex customized extraction pipelines. And unlike earlier multimodal approaches based on CLIP (Contrastive Language-Image Pre-Training) the new approach also enables retrieval of detailed text passages from normal text documents. In this talk we will present how these methods can be implemented using OpenSearch, how this compares to other implementations, and evaluate how well the resulting multimodal RAG system performs in terms of search effectiveness and efficiency.

Machine learning can feel complex, but what if it didn’t have to be? OpenSearch takes the power of ML and puts it right at your fingertips—no extra headaches required. From detecting anomalies in real time to delivering hyper-relevant search results with vector data, OpenSearch is the ultimate partner for turning your data into actionable insights.

In this session, we’ll explore how OpenSearch effortlessly integrates machine learning into your workflows, whether through k-NN search, anomaly detection, or real-time predictions. We’ll also showcase innovative use cases like fraud detection, personalized recommendations, and predictive analytics—all powered by OpenSearch.

If you’re looking to make machine learning accessible, scalable, and downright useful, this talk will inspire you to rethink how you approach data and decision-making with OpenSearch.

In the realm of search, Machine Learning plays a pivotal role in enhancing the user experience throughout the entire lifecycle, from ingesting documents to delivering highly relevant results for user queries. This session will showcase various ML integrations of OpenSearch tailored to optimize outcomes for user queries in a retail scenario, accompanied by a live demonstration. We will explore cutting-edge techniques such as query understanding and rewriting using Large Language Models (LLMs), document enrichment, sparse, dense, and hybrid retrievers, as well as contextual re-ranking of results. As a highlight, you will learn how LLM agents and OpenSearch’s ML Search types can be integrated to build a AI shopping assistant that offers a conversational shopping experience for the users.

Scaling vector search to a billion vectors isn’t just a technical challenge—it’s a high-stakes balancing act of latency, cost, and quality. In this talk, I’ll share practical lessons learned from building a billion-scale system in OpenSearch, covering model selection, text processing, hardware essentials, and embedding strategies. But that’s not all: I’ll expose the tips, tricks, and gotchas you’ll need to navigate OpenSearch’s quirks without losing your sanity (or your wallet). Whether you’re scaling up or just getting started, this session will equip you to tame your vectors and succeed at scale.