JSON Schema Builder

Hand-writing schemas is tedious and error-prone; the examples already contain the structural truth, they just need aggregation. Inferring nullability across N examples (rather than one) prevents the classic mistake of typing a field as string when 3% of your real data has it as null. Pairing inference with LLM-generated descriptions turns a dry schema into documentation that explains itself to downstream consumers.

1) Parse every example file, recursing into nested objects/arrays. 2) For each field path, collect the set of observed types. 3) If more than one type appears and null is one of them, emit oneOf: [<T>, null]. 4) For string fields with ≤ 12 unique values across examples, emit an enum. 5) For numeric fields, emit minimum/maximum pulled from observed range. 6) LLM pass: generate a one-line description per field name using the field path + 3 sample values. 7) Validate the emitted schema against the original examples; any failure triggers a second pass that widens the constraint.