Flutter Generative UI (GenUI) Best Practices

Goal

Implement a dynamic user interface generated by AI using the genui package. This is the recommended approach when an application needs to build reactive screens or components directly from AI responses without hardcoding the layout.

Instructions

The genui package orchestrates a conversational loop between the user, the AI (via ContentGenerator), and the Flutter UI (via A2uiMessageProcessor and DataModel).

1. Project Setup

Add the necessary packages to your pubspec.yaml:

flutter pub add genui
# For Gemini integration via Firebase:
flutter pub add genui_firebase_ai

If targeting macOS/iOS, enable outbound network requests in entitlements:

<key>com.apple.security.network.client</key>
<true/>

2. Core Integration Structure

To connect your Flutter application to the AI agent, you need to instantiate three main components:

1. A2uiMessageProcessor: Handles the UI side, translating AI messages into UI actions. Supply it with a Catalog of widgets the AI is allowed to use.
2. ContentGenerator: Handles communication with the AI. Supply it with system instructions and tools.
3. GenUiConversation: The facade coordinating the processor and generator.

Example Setup in a StatefulWidget:

import 'package:flutter/material.dart';
import 'package:genui/genui.dart';
import 'package:genui_firebase_ai/genui_firebase_ai.dart';

class GenUIPage extends StatefulWidget {
  const GenUIPage({super.key});

  @override
  State<GenUIPage> createState() => _GenUIPageState();
}

class _GenUIPageState extends State<GenUIPage> {
  late final A2uiMessageProcessor _a2uiMessageProcessor;
  late final GenUiConversation _genUiConversation;

  @override
  void initState() {
    super.initState();

    // 1. Initialize Message Processor with predefined catalog widgets
    _a2uiMessageProcessor = A2uiMessageProcessor(
      catalogs: [CoreCatalogItems.asCatalog()],
    );

    // 2. Initialize Content Generator (e.g., Firebase AI / Gemini)
    final contentGenerator = FirebaseAiContentGenerator(
      catalog: CoreCatalogItems.asCatalog(),
      systemInstruction: '''
        You are an AI assistant. You must generate UI by responding with JSON
        matching the provided widget schemas. Always output valid UI structures.
      ''',
      tools: _a2uiMessageProcessor.getTools(),
    );

    // 3. Orchestrate the Conversation
    _genUiConversation = GenUiConversation(
      a2uiMessageProcessor: _a2uiMessageProcessor,
      contentGenerator: contentGenerator,
      // Implement your handlers for rendering generated surfaces
      onSurfaceAdded: (surfaceId) { /* Handle surface creation */ },
      onSurfaceDeleted: (surfaceId) { /* Handle surface deletion */ },
      onTextResponse: (text) { /* Optional: show text explicitly */ },
      onError: (error) { /* Handle generation errors */ },
    );
  }

  @override
  void dispose() {
    _genUiConversation.dispose();
    super.dispose();
  }

  @override
  Widget build(BuildContext context) {
    // UI Implementation interacting with _genUiConversation.sendRequest()
    return const Scaffold();
  }
}

3. Data Model & Data Binding

A core philosophy of genui is utilizing a central DataModel. Widgets don't manage their own dynamic state; instead, they bind to paths within the DataModel.

When creating custom widgets to add to your Catalog, ensure they react to the data context:

// The AI will generate JSON like:
// { "Text": { "text": { "literalString": "Hello" } } }
// OR
// { "Text": { "text": { "path": "/user/welcome_message" } } }

• literalString: Static data directly from the AI.
• path: A reference to a location in the DataModel. Updates to this path automatically rebuild any widgets bound to it. Input widgets (like TextField) should directly update this DataModel.

4. Customizing the Catalog

To extend GenUI beyond the CoreCatalogItems, define your own CatalogItems and provide them to both the A2uiMessageProcessor and the ContentGenerator. This tells the AI what Custom Widgets are available and what parameters they accept.

Constraints

• Architecture Scope: Do not use genui for static screens or standard layouts. It should be reserved exclusively for features needing generative or highly dynamic, AI-driven UI presentation.
• State Management: Within GenUI components, prefer the unified DataModel reactivity over local setState for data shared with the AI.
• Error Handling: Always ensure onError in GenUiConversation is handled appropriately, as AI responses can sometimes fail parsing or violate catalog schemas.