Point RDF Studio at a database schema and some sample rows, and it builds a first model for you: an OWL ontology, SHACL shapes that check your data, controlled vocabularies, and a record of which class and property came from which table and column. This guide says exactly what it takes from each thing you upload — so you know what to give it, and what you will still have to decide.
The rule behind all of it: what can be computed is computed, and only meaning is left to AI. A column declared NOT NULL does not need a language model to tell anyone it is required.
NOT NULL
Upload the schema script your database produces — a .sql file. Nothing in this list involves AI.
.sql
varchar(40)
PRIMARY KEY
UNIQUE
FOREIGN KEY
REFERENCES
CHECK (UnitPrice >= 0)
Types are read for SQL Server, PostgreSQL, MySQL and Oracle scripts. A type the generator does not recognise is kept as text and the model says so — it is never quietly guessed. One SQL Server detail is handled for you: there, timestamp is a row-version counter, not a date, and it is typed as binary.
timestamp
A script with every table in one batch, or split by GO, reads the same. One exception for now: if your script has no GO lines — the usual shape of a PostgreSQL, MySQL or Oracle export — put the views and stored procedures in a .sql file of their own. In the same file as the tables they currently hide the tables, and the Extract step shows 0 tables. Keys and checks written on the column line are read as well as the ones declared at the end of the table.
GO
Views and stored procedures show how the data is actually used, so include them.
JOIN … ON a.x = b.y AND …
FROM a, b WHERE a.x = b.y
A view does not yet become a class of its own.
Add rows for each table as .csv, .xlsx, .json or .parquet. Up to 100 rows per file are read by default. Everything learned from samples is marked inferred, so you can tell it from what your schema declared.
.csv
.xlsx
.json
.parquet
AI is used for meaning: naming things well, describing them from your documents, and reading a classification out of prose. Everything it proposes is a suggestion you review. If you upload only documents, you get an ontology but no validation shapes and no mappings, because those come from declared columns and constraints.
Then read the Extract step's summary before you continue: how many tables, keys, joins, views and procedures were read, and what could not be read.