MTS launches Metan: an AI assistant for finding data in a catalog of 500,000 tables

MTS has built Metan, an intelligent assistant that helps analysts find the data they need in a corporate catalog of 500,000 tables simply by asking a question in Russian.

A catalog that doesn't answer questions

MTS has accumulated enormous volumes of data. The company's data catalog contains more than 500,000 registered tables. Hundreds of specialists work with this data every day: analysts, data engineers and machine learning specialists who build data marts for ML models.

But here is the problem: the catalog works well only if you already know what you are looking for. Need a table with customer information? You type a query, and the catalog returns field descriptions, the owner and the schema. But if you face a business question such as "which tables contain information about customers in the Moscow region for the last quarter?", the catalog won't help. The analyst has to work it out alone: read documentation, consult colleagues from different teams, study the links between sources. This can take hours or even days.

Why a neural network can't do it alone

It seems logical that a neural network could solve this task: just ask it where to find the data. But there is a critical problem: AI relies only on the information it is given. If the metadata contains no explicit information about the links between tables, about how data from one source relates to another, or about what each term actually means in the company's context, even the most advanced model will stall.

MTS, however, has a rare combination of two circumstances. Over 15+ years of systematic data governance work, the company has built serious expertise in describing metadata properly, structuring links and documenting business processes in data. At the same time, LLMs have appeared that can work with such structured semantic layers.

The Metan system: data governance meets AI

Metan is a name that blends the Russian words for "metadata" and "analytics". It is a pilot system that MTS is testing right now. The system acts as an intelligent interface to corporate data: you ask a question in human language, the system works out which tables and sources you need, and gives you an answer.

It works because MTS metadata is not just a set of descriptions. It is a full knowledge graph:

• Tables linked to their descriptions and field structure
• A glossary of the company's business terms
• Links between tables and data sources
• Information about responsible parties and owners
• The history of dependencies and links between data marts

A semantic layer that LLMs can understand sits on top of this graph. The model can "see" not just words and descriptions but meanings and the relationships between them.

What this means for business

Metan is not the first attempt to automate data discovery in large companies. But it is one of the first systems that genuinely works, because it is built not on pure AI but on a combination of costly data governance work and the modern capabilities of LLMs.

For companies like MTS, this means faster analyst work: less time spent searching for sources and coordinating with colleagues, more time for real analysis. For the industry as a whole, it is a signal: agents work better not with chaotic data but with well-organized, well-described data. Order and AI are not enemies but partners.