Skip to main content

Natural Language Rule Builder

The NL Rule Builder at /governance/nl-rules lets you write data-quality rules in plain English. The page calls a Foundation Model endpoint to parse your description into a structured rule (SQL + JSON metadata) that you can review, edit, and save.

Why use it​

DQ rules in JSON or SQL are tedious to author. "Email addresses in the customers table must match a valid pattern and not be null" is faster to write than the equivalent SQL WHERE clause. The model converts intent to structure; you stay in control of what gets saved.

Three modes​

1. Single rule​

Type a rule in natural language plus the table FQN, click Parse Rule:

Description: Total order amount must be between 0 and 1,000,000
Table FQN:   prod_warehouse.sales.orders

The page calls:

POST /nl-rules/from-natural-language
{
  "text": "Total order amount must be between 0 and 1,000,000",
  "table_fqn": "prod_warehouse.sales.orders"
}

Returns:

  • Parsed rule JSON (column, check type, parameters)
  • A confidence badge (high / medium / low)
  • Generated SQL preview

Click Save to persist (POST /governance/dq/rules).

2. Batch​

Textarea takes one rule per line. Parse Batch runs them all and shows a per-rule result with skip option.

3. Explain​

Paste an existing rule JSON and click Explain. The model returns a plain-English description — useful when you've inherited a rule pack and need to document it.

Confidence​

The badge indicates how sure the parser is about its mapping:

  • High — column / operator / values resolved unambiguously
  • Medium — review the parsed JSON; one or more fields inferred
  • Low — likely needs editing before saving (the Save button warns)

Always review parsed JSON before saving low-confidence rules.

Where parsed rules go​

Saved rules appear in the Data Quality Rules Engine and run on the same schedule as hand-written rules.