OpenSearchCon Europe 2026

OpenSearch's ML Commons plugin enables deploying embedding models directly on your cluster—yet many developers still rely on external services. This session demonstrates building a complete RAG system using OpenSearch as both vector store and ML inference engine.

We'll cover deploying sentence-transformer models via ML Commons, processing PDFs, generating embeddings within OpenSearch, configuring knn_vector indices, and implementing semantic search. The highlight: voice search using local Whisper (STT) model creating a fully self-contained system.

What I hope to achieve: Share practical patterns for leveraging OpenSearch's ML capabilities and gather community feedback.

What attendees gain: Step-by-step knowledge to deploy ML models on OpenSearch, build RAG pipelines, and integrate local voice processing—reproducible techniques they can apply immediately.

How this helps the ecosystem: Showcases OpenSearch's native ML features, reducing dependency on external services. Demonstrates OpenSearch as a complete AI platform, not just a search engine.

Target audience: Developers building RAG applications, OpenSearch operators exploring ML capabilities.

Choosing the right ingestion tool for OpenSearch is a critical architectural decision. Should you use the OpenTelemetry Collector or Data Prepper? This session provides a definitive guide to answering that question, helping you design a pipeline that is both powerful and efficient.

This talk moves beyond theory into a practical comparison of design philosophies. We'll contrast the Collector's strength as a universal, lightweight agent with a vast component ecosystem, against Data Prepper's role as a purpose-built pipeline engine, optimized for powerful transformations and accessible stateful processing.

To make this concrete, the session will showcase the setup of a complete hybrid pipeline, walking through the necessary configurations. We will then deconstruct a real-world APM pipeline architecture to show how Collector and Data Prepper fulfill distinct roles in processing data before it reaches OpenSearch.

You'll leave with a clear decision framework, proven architectural patterns, and a blueprint for building your own scalable observability pipelines.

Dense vector retrieval based on single embeddings has become the default approach for semantic search, but it introduces well-known limitations when dealing with long documents, fine-grained relevance signals, or complex queries. Late interaction retrieval models address these issues by allowing queries and documents to interact at a more granular level, typically across token-level or sub-vector representations, while remaining more scalable than full cross-encoders.

This session explores late interaction models from a systems and search-engine perspective, focusing on what it would take to support them efficiently in OpenSearch-based architectures. It discusses indexing and storage strategies for multi-vector representations, candidate generation and pruning techniques, and the role of late interaction scoring as an intermediate stage between first-pass retrieval and full reranking. The talk emphasizes practical trade-offs in latency, memory, and relevance, and outlines how late interaction models could complement existing lexical, neural sparse, and dense vector approaches in production OpenSearch deployments.

Lucene Vector Search Update

There have been a lot of exciting developments in Lucene's vector search implementation recently: so many it's hard to keep track. This talk will survey all the many contributions that have come in over the last year or two, give examples of how to make use the new capabilities, and point the way for future work in this area.

Some of the specific topics may include:

* seeded search (start graph walk from a known point)
* optimistic search (safe pro-rating for efficient multi-segment search)
* Acorn-based search (more efficient application of filters)
* advances in quantization
* efficient indexing using binary partitioning over vector fields
* more efficient merging: finally, we can re-use information from graphs in existing segments and don't have to start from scratch
* proposed integrations: FAISS, CUVS, DiskANN

We'll dive more deeply into some selected topics, but the general idea is to convey the breadth of activity and the diversity of contributors.

As OpenSearch clusters grow beyond 10–20 nodes, centralized monitoring becomes a bottleneck: single points of failure, API overload, and higher latency. Traditional approaches don’t scale.

This session presents a production-ready distributed monitoring architecture using Apache NiFi and MiNiFi. MiNiFi agents on each node collect local metrics via a custom NodeStatsProcessor (CPU, heap, JVM, I/O, thread pools), while central NiFi collectors aggregate and deduplicate cluster-wide metrics using a custom ClusterStatsProcessor and forward them to OpenSearch.

Results include linear scalability, sub-millisecond node metrics, HA, and minimal overhead. Key insights: separating local vs cluster-wide collection, deployment patterns (bare metal, VMs, containers), HA strategies with multiple NiFi collectors, and lessons from production clusters of 10–50+ nodes processing millions of metrics daily with 99.9% reliability.

Ideal for OpenSearch operators managing 10+ nodes, platform engineers building observability pipelines, and anyone hitting centralized monitoring limits. Walk away with a distributed architecture you can implement immediately.

Search relevance tuning is one of the hardest—and most expensive—problems in modern search systems. It demands deep expertise in Lucene scoring, complex query DSLs, and iterative manual testing. While tools like OpenSearch User Behavior Insights (UBI) and the Search Relevance Workbench provide the data and the environment for improvement, the leap from "analyzing data" to "deploying a fix" remains a significant hurdle for many search engineers.

In this session, we introduce Agentic Relevance Tuning (ART)—an end-to-end framework that uses LLM-powered agents to automate the full relevance lifecycle in OpenSearch. ART replaces time consuming manual workflows with a coordinated system of specialized agents that continuously detect relevance issues, hypothesize improvements, and orchestrate offline and online evaluation.

By combining OpenSearch’s relevance tooling with the reasoning and orchestration capabilities of agents, ART transforms relevance tuning into a faster, more accurate, and more accessible process—enabling both beginner and expert search engineers to ship better search experiences to their customers.

GenAI hallucinations plague production apps, but OpenSearch's vector search and hybrid retrieval turn unreliable LLMs into precision engines. This talk unveils battle-tested RAG pipelines that slash errors by 80%+, drawing from semantic search and real-time analytics use cases.

Discover hands-on strategies:
- Ingesting embeddings at scale with neural ranking and k-NN plugins for sub-second queries.
- Hybrid BM25 + vector fusion to ground responses in enterprise knowledge bases and logs.
- Live demos: RAG over petabyte traces, chunking optimizations, and eval frameworks benchmarking recall.

Packed with code snippets, failure autopsies, and deployment blueprints for Kubernetes. Ideal for ML engineers and architects building trustworthy GenAI, escape the hallucination trap and deploy survival-ready search today!

Running OpenSearch in Kubernetes sounds simple — until you try to do it at scale, with high availability, zero downtime, and evolving workloads. In this talk, we draw on years of production experience operating OpenSearch clusters for SAP Cloud Logging service. We aim to share the key lessons learned and design patterns from building our Kubernetes operator.

We’ll dive into how we approached seamless upgrades, built tailored monitoring, and improved performance when interacting with OpenSearch APIs. Even if we adapt solutions to fit internal needs, the lessons learned are valuable for the broader OpenSearch and Kubernetes community.

In the second half, we turn our focus to the open-source OpenSearch Kubernetes Operator. We’ll explore its current state, the challenges the community is addressing and efforts to improve its reliability. We’ll also discuss how this open-source offering enables organizations to seamlessly scale and manage OpenSearch clusters on Kubernetes, and highlight opportunities for collaboration within the project.