Drafting a Domain Glossary with a Model

Writing descriptions for ten tables is manageable. With hundreds, much of the work becomes investigating abbreviations and relationships. A model can prepare a first draft of the glossary for someone familiar with the data to review.

The inputs are table names, columns, available documentation, and, where needed, a small sample of values. The output is a short description and a few terms a user might put in a question.

Give it enough context to make a useful proposal

Names and birthdates in H_OSOBA suggest person records. They do not establish that every person is an employee. That distinction shows the limit of inferring business meaning from structure.

Bound the input: selected columns, a few rows, and truncated long values. The original design uses examples such as ten sample rows; treat that as a starting point to test. If documentation and schema are sufficient, there is no need to include actual values.

Review the output

One common failure is returning column names as aliases. TITUL_PRED_KOD does not help someone asking about an academic title. An automated check can flag that output and request a revision with a specific explanation.

Other errors need domain knowledge. The model may confuse contracts with people, or historical records with current status. A review screen should put the proposed description beside its source material and a question the table needs to support.

Start human review with tables used in real questions. For the rest, keep the distinction between a generated draft and an approved entry visible.

Make a large run resumable

Process the schema in batches, limit concurrent calls, and save completed results as you go. After an API failure, resume with unfinished tables. A rerun should also preserve manual corrections.

Track the source and review status of each entry. Store approved versions with revision history and check them against a set of questions. This makes it possible to expand the glossary without losing the knowledge added during earlier reviews.