A trusted AI context layer for Seznam.cz

The Client

Seznam.cz is the Czech internet leader, one of the very few markets in the world where a homegrown search engine and web portal competes head-to-head with global players. Founded in 1996, it is both a media house and a deeply technical engineering company, operating search, news, maps, e-mail and advertising at national scale for millions of daily users.

With an engineering-first culture and a large in-house tech team, Seznam.cz generates huge volumes of data every day. As it invests in AI and data-driven decision-making, making that data discoverable and trustworthy (for people and AI agents alike) became a strategic priority.

​

The Problem

Seznam.cz set out to make its data answerable in plain language, letting people across the company ask business questions and get a trustworthy answer without writing SQL or knowing where the data physically lives. The vehicle is an internal conversational analytics tool.

But a conversational analytics tool is only as good as the context it can reach. Ask it a seemingly simple question such as “when are there the most visits on Seznam.cz?” and, without a shared definition, the tool simply guesses: it writes ad-hoc SQL and counts something, with no way of knowing what a visit actually means inside Seznam.cz.

A second problem reinforced this: teams sometimes reported different numbers for the same metric, with no easy way to trace where a figure came from or which one was right. And data kept multiplying: for almost every new question or report, yet another dataset was created, instead of reusing what already existed. Seznam.cz’s analytics stack relies heavily on Keboola, Snowflake and Tableau, but there was no end-to-end view of how data flowed between them.

Seznam.cz had piloted an open-source tool, OpenMetadata, but running it in-house proved costly and slow, with significant hidden effort and weak support for the two things that mattered most: lineage and data products.

​

“Processing all of our data centrally, in a single team, had become unsustainable - a real bottleneck. So we moved towards a data-mesh model, with dedicated teams across Seznam’s domains: maps, shopping, advertising, content services and many others. With our analytics data spread across more than 30 Keboola projects, we needed a tool that could map that landscape in real depth - and a shared environment where every team can see which data products already exist, who owns them, and who to talk to when one needs extending. We had tried to build that on OpenMetadata, but that path proved endless. The Dawiso pilot showed real value within a few weeks - its ready-made connectors let us scan Keboola, Snowflake and Tableau almost immediately.”

Jakub Trefilík, Senior Product Manager, Seznam.cz

​

The Solution

Dawiso became the single context layer that both people and AI agents query for one authoritative answer. The work has grown organically, use case by use case, with the AI context layer at its core.

The centrepiece is Dawiso’s MCP server. The conversational analytics tool connects to Dawiso through the Model Context Protocol, so before it composes a query it can ask Dawiso for the company’s own definitions. A visit, for example, has a precise definition in Dawiso (activity from one person on one device within a set time window of a few hours), so refreshing the homepage several times still counts as a single visit. Grounded in that definition, the tool builds correct SQL and returns an answer the business can trust.

The same context layer points the tool to Seznam’s data products. Asked something like “what was the most-read article on Seznam Zprávy yesterday?”, it first has to know what “most-read” means (unique readers? page views?) and where “yesterday” begins and ends, and only then which data to use. Dawiso supplies those definitions and points the tool to the right data product and table IDs, instead of leaving it to hunt blindly across thousands of tables.

Just as important is how Seznam.cz now treats its data. Dawiso handles the design and registry of data products, so needs are consolidated instead of spawning a new dataset every time: when a new requirement appears, an existing data product is extended rather than duplicated. Each one is described, owned and ready to use: no duplication, clear ownership, faster delivery.

The same context also speeds up how new analytics get built. Once a data product and its underlying tables are documented and linked in Dawiso, a semantic view can be generated straight from that context, for example in Claude or Cursor through the MCP server. And because everything is linked, the view knows its lineage: if a metric is already pre-computed in a Tableau dashboard downstream, it can reuse the result instead of burning compute to recalculate it.

To restore trust in reporting, Dawiso built end-to-end data lineage across Keboola, Snowflake and Tableau. Seznam.cz can now see how data flows from Keboola through Snowflake and its transformation layers all the way to Tableau reports, and run impact analysis: if a table changes, exactly which reports downstream are affected. As a highly technical organisation, it also got markdown-native authoring, the natural editor for its engineers and a token-efficient format for AI, which made adoption easy.

Seznam.cz still keeps OpenMetadata for technical work; Dawiso adds the business layer on top. A new employee can start from a plain business term, search for it, and immediately see where the relevant data and dashboards live, by clicking through the web catalog or simply asking in natural language.

Not every source is covered out of the box, and that’s fine. Heavy processing still happens upstream on on-prem Airflow clusters before data reaches Snowflake, tools such as GoodData surfaced in the pilot, and as adoption deepens new needs appear: Airflow connectivity is in PoC, MLOps pipelines further out. Because Dawiso is open and customisable, these are a question of prioritisation, not feasibility: the team has even built its own Agent Catalog on top of it.

​

Main features of the Dawiso solution for Seznam.cz

AI context layer over MCP

Dawiso’s MCP server feeds the conversational analytics tool with authoritative business context (definitions, ownership and the right data sources), so AI answers are accurate and consistent rather than guessed.

​

Governed data products

Scattered tables are consolidated into described, owned, ready-to-use data products covering areas such as audience affinity, socio-economic data and homepage traffic, each with context on what the data is and how to use it. Needs are met by extending existing products rather than spawning new datasets: no duplication, clear ownership, faster delivery.

​

Semantic views generated over MCP

From a data product and its linked sources, Dawiso’s MCP server helps generate semantic views, for example in Claude or Cursor, wired into lineage so they reuse what already exists instead of recomputing it.

​

End-to-end data lineage

A complete, interactive view of how data moves from Keboola through Snowflake and its transformation layers into Tableau reports, enabling impact analysis that Tableau alone cannot provide.

​

Markdown-native, AI-ready documentation

Native markdown authoring matched the team’s engineering culture and keeps documentation lightweight and token-efficient for AI, a clear differentiator delivered quickly.

​

The Impact

Dawiso gave Seznam.cz one place where business meaning, technical metadata and AI consumption meet. The conversational analytics tool now answers grounded in the company’s own definitions, and knows where to find the right data, so the same question returns the same trusted number.

For the teams working across the analytics stack the change is already tangible: end-to-end lineage across Keboola, Snowflake and Tableau makes impact analysis routine, so they know exactly where data comes from and what a change will affect. Data products give the wider organisation clear ownership and discovery, consolidating what used to be scattered, duplicated datasets and shortening the time it takes to find and use the right data.

Just as importantly, the platform is raising data culture and trust across the company. With the foundation in place (context layer, lineage and data products), Seznam.cz is now preparing a broader rollout to business users, building on, rather than replacing, what its teams already do.

​

“What I value most is that Dawiso isn’t just a tool - it brought the knowledge and guidance we were missing. Before Dawiso, ‘data product’ was more of a buzzword for us; now it’s a concrete, owned entity our teams actually work with. Our account manager, František, meets with us regularly and genuinely follows where we’re heading - they even taught us to make smaller edits to our packages ourselves. And when we said we wanted an Agent Catalog, the new interface was ready in a couple of days; planning that into a sprint anywhere else would have taken weeks.”

Jakub Trefilík, Senior Product Manager, Seznam.cz

​

Next phase: governing AI agents in Snowflake Cortex

Currently piloting, the next step will extend the context layer to Seznam’s AI agents themselves. Dawiso will scan Snowflake’s Cortex agents and the semantic views they read from, and will build the lineage automatically: each “talk-to-your-data” agent → its semantic views → the underlying tables. A business user will be able to find the right agent, see exactly which data it reads, and trust its answers.

It will also work the other way. Since a semantic view is essentially a data product, Dawiso will generate governed semantic views from its data products and push them into Snowflake via Open Semantic Interchange (OSI), so teams, even business users, will be able to spin up new, governed agents (sales, HR, customer care...) straight from the catalog, with full governance over what each one reads.