Context is King: Building AI Agents on Real Data
Context is king when it comes to building meaningful AI agents and everyone knows it, but very few are doing it right at scale.
This is a meetup for Helsinki-based developers and builders for sharing real experiences about working with problems related to semantic layer, metadata, governance and more.
Agenda:
17.00: Doors open
17.15: A Brief History of Context: the Evolution of Metadata (Juha Korpela, Datakor)
17.40: Why Your Data Catalog is Your AI's Brain (Stan Dmitriev, Aiven)
18.05: Break
18.15: Building Reliable Data Agents with a Semantic Context Layer (Karolus Sariola, Flow AI)
18.40: Context and RAG in Practice (Edward Kim, Confidential Mind)
19.05: Context Management for Agentic RAG (Johan Jern, Realm)
19.30: Open discussion
Context:
A Brief History of Context: the Evolution of Metadata (Datakor)
To understand data, we need information about the data: metadata. This is the context around our data that AI solutions are so hungry for. But while the demand for metadata has now exploded, there's nothing new in the concept itself. We've always produced, managed, and utilized various kinds of metadata.
In this session, we'll have a look at the big picture of metadata management and assess how we've got where we are now. Managing context means managing technical metadata, semantic metadata, and governance metadata: these are not new problems. When you understand the big picture, you'll see that not all solutions are new, either - instead of reinventing, we are building on what has come before.
Why Your Data Catalog is Your AI's Brain (Aiven)
Data catalogs are no longer just for compliance - they are the essential context layer for AI. While LLMs are powerful, they fail without the "tribal knowledge" found in metadata.
This session explores how to transform your catalog from a static repository into an active nervous system for AI agents.
We’ll dive into:
Metadata Shift-Left: Capturing context at the source.
MCP as an Interface: Using the Model Context Protocol to let AI reason with data.
Automated Enrichment: Using AI to finally solve the "empty catalog" problem.We'll conclude with some hot takes on why modern catalogs are broken and what it takes to actually fix the metadata loop
Building Reliable Data Agents with a Semantic Context Layer (Flow AI)
AI agents working on real data often fail not because of model quality, but because meaning is unclear, inconsistent, or changes over time. Dumping schemas, tables, or documents into model context—or even generic RAG—doesn’t scale when agents need to reason over complex, multi-tenant data environments.
In this session, we’ll share how we think about making data systems agent-ready through a semantic context layer and a data agent architecture designed for reliability and adaptability. We’ll discuss how agents can incrementally discover and validate meaning, how semantics can evolve safely as data changes, and how different agents can operate over the same data with different principles.
We’ll also include a short demo from a recent customer case to illustrate how these concepts apply in a real-world setting.
Context and RAG in practice (ConfidentialMind)
RAG systems don’t usually fail because the LLM is “dumb” — they fail because the wrong context gets retrieved and fed into the model. In this talk, we do a walkthrough of a real-world RAG failure and fix it in minutes by making context explicit. We’ll start with a naive semantic retrieval setup that returns a plausible but incorrect answer (e.g., wrong version, wrong region, or outdated policy). Then we’ll improve it step-by-step using three practical “context knobs” used in production:
-Hybrid retrieval (semantic + BM25/keyword) to handle exact terms and reduce noisy matches
-Metadata + scope filtering (e.g., region, doc type, owner, version) to prevent “multiple truths” from colliding
-Retrieval tracing to show why specific chunks were selected and how to debug RAG
The goal is to give builders a concrete playbook for selecting higher-quality context without adding more slides, more prompts, or more tokens.
Context Management for Agentic RAG (Johan Jern, Realm)
Some queries are hard to solve with "basic" RAG. When questions require multi-step reasoning, full-document understanding (not just chunks), or aggregating many results that match specific criteria, simple retrieve-and-generate pipelines break down, we need agentic RAG. But this added capability comes at a cost: as agents plan, search, read, and iterate, they quickly use up a lot of context, which both degrades answer quality and increases costs and latency.In this session, we’ll focus on how to tackle the context-related challenges with agentic systems focused on retrieval.
We'll cover:
-What to fill the context with: what tools to give the agent and how it should use them.
-How to manage the context: strategies for context truncation.