Flutter iOS local AI recipes

Here are small tasks you can adapt to your Flutter app: summarize a passage, extract appointment fields and classify a note. Start with installation and availability. Each recipe uses ModelMode.local and CloudPolicy.never to keep generation on the device. The snippets were reviewed against the source and were not executed.

Summarize a short passage

Use the complete availability-checked, session-owned example in the English tutorial or Spanish tutorial. Use supplied facts, bound input and output, and keep the manual path available.

Extract a person’s name and an explicitly supplied day

Pass a schema instead of relying on a prompt that merely requests JSON. This function returns a decoded map for application review, or null when the model is unavailable or the expected fields cannot be accepted.

import 'package:cupertino_fundations_models/cupertino_fundations_models.dart';

Future<Map<String, Object?>?> extractAppointment(String text) async {
  final models = CupertinoFoundationModels();
  try {
    final availability = await models.checkAvailability(
      mode: ModelMode.local,
      cloudPolicy: CloudPolicy.never,
      localeIdentifier: 'en_US',
    );
    if (!availability.isAvailable) return null;

    final response = await models.generateStructured(
      prompt: Prompt.text(text),
      mode: ModelMode.local,
      cloudPolicy: CloudPolicy.never,
      instructions:
          'Extract only the person and day explicitly written in the text. '
          'Use an empty string for a missing field. Do not infer a calendar date.',
      schema: const StructuredSchema.object(
        name: 'Appointment',
        properties: <String, SchemaProperty>{
          'person': SchemaProperty.string(),
          'day': SchemaProperty.string(),
        },
        requiredProperties: <String>['person', 'day'],
      ),
      options: const GenerationOptions(maximumResponseTokens: 120),
    );
    final value = response.structuredValue;
    if (value is! Map<String, Object?>) return null;
    if (value['person'] is! String || value['day'] is! String) return null;
    return value;
  } on FoundationModelsException {
    return null;
  }
}

The facade owns session disposal for generateStructured(). A valid schema does not prove the extracted name/day is correct. Compare with the input and ask the user before creating an appointment. See document extraction for a larger bounded example.

Classify text into a small vocabulary

An enum prevents arbitrary category strings. Keep an other label for text outside the feature’s scope and validate the decoded result independently.

import 'package:cupertino_fundations_models/cupertino_fundations_models.dart';

Future<String?> classifyNote(String text) async {
  final models = CupertinoFoundationModels();
  const categories = <String>['work', 'personal', 'other'];
  try {
    final availability = await models.checkAvailability(
      mode: ModelMode.local,
      cloudPolicy: CloudPolicy.never,
      localeIdentifier: 'en_US',
    );
    if (!availability.isAvailable) return null;

    final response = await models.generateStructured(
      prompt: Prompt.text(text),
      mode: ModelMode.local,
      cloudPolicy: CloudPolicy.never,
      instructions:
          'Classify the note as work, personal or other. '
          'Choose other if the text is ambiguous or unrelated.',
      schema: const StructuredSchema.object(
        name: 'NoteCategory',
        properties: <String, SchemaProperty>{
          'category': SchemaProperty.string(enumValues: categories),
        },
        requiredProperties: <String>['category'],
      ),
      options: const GenerationOptions(maximumResponseTokens: 60),
    );
    final value = response.structuredValue;
    if (value is! Map<String, Object?>) return null;
    final category = value['category'];
    if (category is! String || !categories.contains(category)) return null;
    return category;
  } on FoundationModelsException {
    return null;
  }
}

The label is an AI suggestion, not an authorization or irreversible decision. Allow correction and preserve a deterministic default.

Add streaming, tools or speech

README streaming example: cumulative snapshots, completion, failure and finally disposal.
Guided streaming: only terminal JSON is final.
Tool calling: validate authorization in the app’s tool.
Speech transcription: separate privacy, permissions and assets.
Example source: existing Flutter integration.

Do not run multiple requests on one session. Await cancellation and native cleanup before reuse. Use troubleshooting for typed failure recovery; do not hide failure behind generated success text.