typed_llm_generator 0.2.3

build_runner code generator that turns @LlmSchema-annotated classes into JSON Schemas and validated-JSON factories for the typed_llm package.

example/README.md

typed_llm_generator example

This package is a build_runner code generator — you never call it from Dart code, so this example shows the input it consumes and the output it produces. For a runnable app, see packages/example in the repository, which wires this generator up against a real provider.

1. Depend on it

dependencies:
  typed_llm: ^0.2.0

dev_dependencies:
  build_runner: ^2.4.0
  typed_llm_generator: ^0.2.0

Code generation requires Dart 3.9 or newer. The typed_llm runtime package itself supports Dart 3.4+.

2. Annotate a class

Add the part directive and annotate. That's the whole input — no factory or fromJson boilerplate to write:

// lib/invoice.dart
import 'package:typed_llm/typed_llm.dart';

part 'invoice.g.dart';

@LlmSchema()
class Invoice {
  Invoice({
    @LlmField(description: 'The legal name of the vendor issuing the invoice')
    required this.vendorName,
    required this.totalAmount,
    required this.dueDate,
  });

  final String vendorName;
  final double totalAmount;
  final DateTime dueDate;
}

@LlmField(description:) is optional, but the model sees those descriptions in the schema — they measurably improve extraction accuracy on ambiguous field names.

3. Generate

dart run build_runner build

4. What you get

Three top-level members in lib/invoice.g.dart:

/// The [LlmType] for [Invoice] — pass this to `Extractor.extract`.
const LlmType<Invoice> $Invoice = LlmType<Invoice>(
  name: 'Invoice',
  schema: InvoiceSchema,
  fromJson: _$InvoiceFromValidatedJson,
);

/// The raw JSON Schema describing [Invoice].
const Map<String, dynamic> InvoiceSchema = {
  'type': 'object',
  'properties': {
    'vendorName': {
      'type': 'string',
      'description': 'The legal name of the vendor issuing the invoice',
    },
    'totalAmount': {'type': 'number'},
    'dueDate': {'type': 'string', 'format': 'date-time'},
  },
  'required': ['vendorName', 'totalAmount', 'dueDate'],
  'additionalProperties': false,
};

Invoice _$InvoiceFromValidatedJson(Map<String, dynamic> json) {
  return Invoice(
    vendorName: json['vendorName'] as String,
    totalAmount: (json['totalAmount'] as num).toDouble(),
    dueDate: DateTime.parse(json['dueDate'] as String),
  );
}

5. Use it

$Invoice carries the schema and the parser together, so extract infers the return type and a mismatched pair cannot be expressed:

final invoice = await extractor.extract($Invoice, prompt: 'Extract the ...');

InvoiceSchema remains available if you want to send or inspect the raw JSON Schema yourself.

Supported field types

String, int, double, num, bool, DateTime (ISO-8601 date-time string), enums (string enum), nullable fields, List<T> of any supported type, and other @LlmSchema() classes — nested classes are resolved recursively and inlined into both the schema and the parser. freezed classes work too: the generator reads constructor parameters from the redirecting const factory the same way it reads a plain generative constructor.

Anything else is a build-time error naming the offending field and type, rather than a silently missing property in the schema.