Teachings Generator

The current public Text2SQL package does not export a TeachingsGenerator class or a text2sql.teach() convenience method. If you want schema-aware fragment bundles today, use one of the supported patterns below.

Option 1: Build a Manual Fragment Bundle

Create a fragment array with the domain terms, hints, guardrails, and examples you want to keep stable:

import { example, guardrail, hint, term } from '@deepagents/context';

const teachings = [
  term('NPL', 'non-performing loan - loan past due 90+ days'),
  hint("Always exclude accounts where account_type = 'test'"),
  guardrail({
    rule: 'Never expose full account numbers',
    reason: 'Sensitive financial data',
    action: 'Return masked or aggregated results instead',
  }),
  example({
    question: 'show me delinquent loans by region',
    answer: `SELECT region, COUNT(*) AS delinquent_loans
FROM loans
WHERE days_past_due >= 30
GROUP BY region
ORDER BY delinquent_loans DESC`,
  }),
];

Set this array on your ContextEngine for chat(), or pass it to the lower-level toSql({ fragments }) helper for direct SQL generation.

Option 2: Feed Teachings into SchemaSynthesizer

If you are generating synthetic training data, SchemaSynthesizer accepts a teachings array directly:

import { SchemaSynthesizer } from '@deepagents/text2sql/synthesis';

const pairs = await new SchemaSynthesizer(adapter, {
  count: 10,
  complexity: 'medium',
  teachings,
}).toPairs();

Suggested Workflow

1. Start with a small manual bundle based on real product terminology.
2. Test it against chat() or the low-level toSql({ fragments }) helper.
3. Refine vague or noisy teachings before reusing them in synthesis pipelines.
4. Promote stable bundles into recipe files once they prove useful.

What Changed

Older docs referenced a TeachingsGenerator class and text2sql.teach() method. Those surfaces are not part of the current public API, so the supported workflow is manual fragment bundles plus synthesis-time teachings injection.