How We Described 15,000 Tables in Six Months Instead of 500 in a Year – and Stopped Writing Documentation by Hand

Yandex faced a massive data documentation challenge: out of 40,000 filtered tables, only 500 got manual descriptions in a year. Roman Gridnev, a technical manager, built a Python robot that uses LLMs and internal data sources (data paths, schemas, sample data, Wiki glossaries, graph neighbors from Memgraph, folder neighbors) to generate draft descriptions. The robot first gathers objective data, then sends it to an LLM to produce a title and description, avoiding a blank page. Descriptions are loaded into the Data Catalog, where humans can review them. In six months, 15,000 tables were described, saving about five years of analyst time. Quality was measured by surveys and semantic comparison (92% semantic match, 71% of fields accepted unchanged). Future plans include connecting to regular data deliveries and per-column description propagation.