flutter_ai_assistant

KO | EN

gitlite — search

#hacktoberfest #react #golang #android #rust #deep-learning #docker #macos #nodejs #windows #cli #pytorch #openai #mcp

flutter_ai_assistant

(★ 21)

Drop-in AI assistant for Flutter apps. Reads UI via Semantics tree, executes multi-step tasks autonomously, and works with Gemini, Claude, or OpenAI.

flutter_ai_assistant 최신버젼 다운로드

최종 버전 다운로드 (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# flutter_ai_assistant

A drop-in AI assistant for Flutter apps. Understands your UI through the Semantics tree, executes multi-step tasks autonomously, and works with any LLM provider — Gemini (via Firebase AI Logic), Claude, or OpenAI.

**One widget. Full app control. Zero hardcoding.**

```dart
AiAssistant(
  config: AiAssistantConfig(
    provider: FirebaseAiProvider(
      firebaseAi: FirebaseAI.googleAI(),
      model: 'gemini-2.5-flash',
    ),
  ),
  child: MaterialApp(home: HomeScreen()),
)
```

Your users can now say _"order 2 onions from the store"_ and the assistant will navigate to the store, search for onions, add them to cart, adjust the quantity, and proceed to checkout — all autonomously.

---

## Demo

<table>
  <tr>
    <td align="center"><strong>Ordering from a Store</strong></td>
    <td align="center"><strong>Booking a Ride</strong></td>
  </tr>
  <tr>
    <td><img src="https://storage.googleapis.com/myrik-app-assets/rider/store_ai_assistant_edited.gif" width="300" alt="AI assistant ordering from a store"></td>
    <td><img src="https://storage.googleapis.com/myrik-app-assets/rider/ride_ai_assistant_edited.gif" width="300" alt="AI assistant booking a ride"></td>
  </tr>
</table>

---

## Table of Contents

- [How It Works](#how-it-works)
- [Features](#features)
- [Quick Start](#quick-start)
- [Configuration Reference](#configuration-reference)
- [LLM Providers](#llm-providers)
- [Migrating from `GeminiProvider`](#migrating-from-geminiprovider)
- [Custom Tools](#custom-tools)
- [App Manifest (Code Generation)](#app-manifest-code-generation)
- [Voice I/O](#voice-io)
- [Analytics Events](#analytics-events)
- [Rich Chat Content](#rich-chat-content)
- [Architecture](#architecture)
- [API Reference](#api-reference)

---

## How It Works

The assistant doesn't use hardcoded screen coordinates or widget keys. It reads Flutter's **Semantics tree** — the same accessibility layer used by screen readers — to understand what's on screen and interact with it. This means it works with any Flutter app out of the box, regardless of your widget structure.

---

## Features

### Core Intelligence
- **ReAct Agent Loop** — Reason, Act, Observe cycle with automatic verification
- **Multi-provider LLM** — Gemini, Claude, and OpenAI out of the box; bring your own via `LlmProvider` interface
- **Built-in Tools** — tap, type, scroll, navigate, go back, long press, increase/decrease values, ask user, hand off to user, read screen (10 always-on + `hand_off_to_user` when `confirmDestructiveActions` is true)
- **Custom Tools** — register your own business-logic tools (check inventory, call APIs, etc.)
- **Conversation Memory** — multi-turn context with automatic management
- **Circuit Breaker** — escalates after consecutive failures instead of looping forever

### UI Understanding
- **Semantics Tree Walking** — reads every interactive element, label, and state on screen
- **Progressive Screen Knowledge** — remembers previously visited screens for smarter planning
- **Screenshot Support** — optional visual context for chart/image understanding (multimodal)
- **Context Caching** — avoids redundant tree walks with configurable TTL
- **App Manifest** — optional code-generated "building map" of your entire app for instant navigation

### Voice
- **Speech-to-Text** — multi-locale recognition with automatic locale selection
- **Text-to-Speech** — auto-detects Hindi/English and switches voice accordingly
- **Confidence Filtering** — discards low-confidence noise before it reaches the LLM
- **Summary Mode** — speaks only the first sentence of long responses; full text stays in chat

### UI
- **Floating Action Button** — draggable, edge-snapping, with processing indicator and unread badge
- **Chat Overlay** — full-screen chat with animated action feed showing live progress
- **Rich Messages** — text, images, interactive buttons, and cards in chat bubbles
- **Handoff Mode** — for irreversible actions, the overlay clears so the user can tap the final button themselves
- **Response Popup** — compact result card above the FAB after auto-close
- **Suggestion Chips** — configurable quick-start actions in empty state
- **Post-task Chips** — contextual follow-up buttons after task completion
- **Auto-close** — overlay closes after task completion; action results auto-dismiss, info results persist

### Safety

- **Destructive Action Handoff** — purchases, deletions, and payments are handed to the user for the final tap
- **ask_user Guards** — code-level enforcement prevents the LLM from asking unnecessary confirmation questions
- **Verification Passes** — after the agent says "done", the system re-checks the screen to catch premature completion
- **Max Iterations** — hard cap on agent loop steps to prevent runaway execution
- **Processing Timeout** — 3-minute safety timeout on active processing

---

## Quick Start

### 1. Add the dependencies

```yaml
dependencies:
  flutter_ai_assistant: ^0.2.0
  firebase_core: ^4.3.0
  firebase_ai: ^3.11.0
  # Optional but strongly recommended for production:
  # firebase_app_check: ^0.4.0
```

### 2. Initialise Firebase

`FirebaseAiProvider` routes Gemini calls through Firebase AI Logic, so your host app must initialise Firebase before constructing the provider. Most apps already do this for analytics/auth/messaging.

```dart
import 'package:firebase_core/firebase_core.dart';

Future<void> main() async {
  WidgetsFlutterBinding.ensureInitialized();
  await Firebase.initializeApp(
    options: DefaultFirebaseOptions.currentPlatform,
  );
  runApp(const MyApp());
}
```

> Don't have a Firebase project yet? Run `flutterfire configure` (from the FlutterFire CLI) to wire one up.

### 3. Wrap your app and wire the navigator observer

The assistant needs two things: the `AiAssistant` widget wrapping your app, and its `AiNavigatorObserver` added to your `MaterialApp` so it can track route changes.

```dart
import 'package:firebase_ai/firebase_ai.dart';
import 'package:flutter_ai_assistant/flutter_ai_assistant.dart';

class MyApp extends StatelessWidget {
  @override
  Widget build(BuildContext context) {
    return AiAssistant(
      config: AiAssistantConfig(
        provider: FirebaseAiProvider(
          firebaseAi: FirebaseAI.googleAI(),
          model: 'gemini-2.5-flash',
        ),
      ),
      child: Builder(
        builder: (context) {
          final aiCtrl = AiAssistant.read(context);
          return MaterialApp(
            navigatorObservers: [aiCtrl.navigatorObserver],
            home: HomeScreen(),
          );
        },
      ),
    );
  }
}
```

That's it. A floating AI button appears. Users can tap it, type or speak a command, and the assistant executes it.

> **Why the `Builder`?** `AiAssistant.read(context)` requires the `AiAssistant` to be an ancestor in the widget tree. The `Builder` creates a new context below `AiAssistant` so the controller is accessible. Without the navigator observer, the assistant won't know which screen the user is on.

### 4. Add routes and descriptions (recommended)

Telling the assistant about your app's screens makes it dramatically smarter:

```dart
AiAssistant(
  config: AiAssistantConfig(
    provider: FirebaseAiProvider(firebaseAi: FirebaseAI.googleAI()),
    knownRoutes: ['/home', '/store', '/cart', '/profile', '/settings'],
    routeDescriptions: {
      '/home': 'Main dashboard with quick actions',
      '/store': 'Browse and buy products',
      '/cart': 'Shopping cart with checkout',
      '/profile': 'User profile and account settings',
      '/settings': 'App preferences and configuration',
    },
  ),
  child: MaterialApp(...),
)
```

### 5. Add domain knowledge (recommended)

Teach the assistant your app's vocabulary and workflows:

```dart
AiAssistant(
  config: AiAssistantConfig(
    provider: FirebaseAiProvider(firebaseAi: FirebaseAI.googleAI()),

// What your app does (for the LLM's understanding)
    appPurpose:
      'ShopApp is a grocery delivery app. Users browse products, '
      'add to cart, and checkout. "order"/"buy" = full purchase flow. '
      '"cart" = shopping cart. "balance" = wallet screen.',

// Behavioral rules specific to your app
    domainInstructions:
      'QUANTITIES: Tap ADD first (sets qty=1), then tap "+" to increase. '
      '"5 onions" means ADD then "+" 4 times.\n\n'
      '"order X" means COMPLETE the full purchase: add to cart, '
      'go to cart, checkout, and hand off payment.',

// Example flows the LLM should learn from
    fewShotExamples: [
      'User: "order 2 onions"\n'
      'Actions: navigate_to_route("/store") -> set_text("Search", "onion") -> '
      'tap_element("ADD", parentContext: "Onion") -> increase_value("+") -> '
      'navigate_to_route("/cart") -> tap_element("Checkout")\n'
      'Response: "2 onions added and ready for checkout!"',
    ],
  ),
  child: MaterialApp(...),
)
```

---

## Configuration Reference

Every aspect of the assistant is configurable through `AiAssistantConfig`:

### Required

| Parameter | Type | Description |
|-----------|------|-------------|
| `provider` | `LlmProvider` | The LLM provider to use (Gemini, Claude, OpenAI, or custom) |

### App Knowledge

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `knownRoutes` | `List<String>` | `[]` | All named routes in your app (e.g. `['/home', '/store']`) |
| `routeDescriptions` | `Map<String, String>` | `{}` | Human-readable description for each route |
| `appPurpose` | `String?` | `null` | What your app does — domain vocabulary and user intent mapping |
| `domainInstructions` | `String?` | `null` | App-specific behavioral rules injected into the system prompt |
| `fewShotExamples` | `List<String>` | `[]` | Example User -> Actions -> Response flows for the LLM to learn |
| `appManifest` | `AiAppManifest?` | `null` | Rich hierarchical app description (code-generated) |
| `globalContextProvider` | `Future<Map<String, dynamic>> Function()?` | `null` | Callback providing app-level state (user info, cart, etc.) |

### Behavior

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `confirmDestructiveActions` | `bool` | `true` | Hand off irreversible actions (purchases, deletions) to the user |
| `maxAgentIterations` | `int` | `30` | Max reason-act-observe cycles per user message |
| `maxVerificationAttempts` | `int` | `2` | Max post-completion verification passes |
| `contextCacheTtl` | `Duration` | `10 seconds` | How long to cache a screen's semantics snapshot |
| `navigateToRoute` | `Future<void> Function(String)?` | `null` | Custom navigation callback (default: uses NavigatorState) |
| `systemPromptOverride` | `String?` | `null` | Replace the entire built-in system prompt |
| `customTools` | `List<AiTool>` | `[]` | Additional tools the LLM can call |

### Voice

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `voiceEnabled` | `bool` | `true` | Show mic button and enable voice input |
| `enableTts` | `bool` | `true` | Speak responses aloud via TTS |
| `preferredLocales` | `List<String>` | `['en_US']` | Speech recognition locales in priority order |
| `enableHaptics` | `bool` | `true` | Vibrate on mic activation, progress, and completion |

### UI

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `showFloatingButton` | `bool` | `true` | Show the floating AI chat button |
| `fabBottomPadding` | `double` | `72` | Extra bottom padding to clear bottom nav bars |
| `fabDraggable` | `bool` | `true` | Allow dragging the FAB to reposition |
| `autoCloseOnComplete` | `bool` | `true` | Auto-close overlay after task completion |
| `assistantName` | `String` | `'AI Assistant'` | Display name shown in the chat header |
| `initialSuggestions` | `List<AiSuggestionChip>` | `[]` | Quick-start chips in empty chat state |
| `postTaskChipsBuilder` | `PostTaskChipsBuilder?` | `null` | Callback to build follow-up suggestion buttons |

### Screenshots & Debugging

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `enableScreenshots` | `bool` | `false` | Capture screen images for multimodal LLM context |
| `enableLogging` | `bool` | `false` | Write debug logs via `dart:developer` |

### Analytics

| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `onEvent` | `AiEventCallback?` | `null` | Receive structured analytics events for every assistant action |

---

## LLM Providers

The package includes three providers. All share the same `LlmProvider` interface, so switching is a one-line change.

### Gemini via Firebase AI Logic (recommended)

```dart
import 'package:firebase_ai/firebase_ai.dart';

FirebaseAiProvider(
  firebaseAi: FirebaseAI.googleAI(),    // or FirebaseAI.vertexAI(location: 'us-central1')
  model: 'gemini-2.5-flash',            // default
  temperature: 0.2,                      // default
  thinkingConfig: ThinkingConfig.withThinkingBudget(1024),  // optional, Gemini 2.5+
  requestTimeout: Duration(seconds: 45),
)
```

Routes calls through Firebase AI Logic. **Your Gemini API key never ships in the app binary** — it lives on Firebase. Implicit prompt caching on Gemini 2.5+ is automatic. Supports function calling, multi-modal (images), thinking mode, structured output, and streaming.

Three security postures, in order of decreasing strength:

| Posture | Setup | What protects the key |
|---|---|---|
| **Best** — Firebase + App Check (limited-use) | `FirebaseAI.googleAI(appCheck: FirebaseAppCheck.instance, useLimitedUseAppCheckTokens: true)` | Key never on device + per-request platform attestation, single-use 5-min tokens, ready for May 2026 replay protection |
| **Good** — Firebase, no App Check | `FirebaseAI.googleAI()` | Key never on device. Anyone with your Firebase config can technically still invoke; rely on Firebase quotas and usage alerts |
| **Worst (legacy)** — raw API key in app binary | `GeminiProvider(apiKey: '...')` (deprecated) | Nothing — key is extractable from APK/IPA |

To enable App Check, add `firebase_app_check` to your app's pubspec and call `FirebaseAppCheck.instance.activate(...)` once before `runApp(...)`. See [the Firebase App Check docs](https://firebase.google.com/docs/app-check) for per-platform setup (Play Integrity / App Attest / reCAPTCHA Enterprise).

#### Streaming and structured output

`FirebaseAiProvider` exposes two helpers beyond the base `LlmProvider` contract for use outside the agent loop:

```dart
// Streaming for a custom chat UI:
await for (final event in provider.streamMessage(
  messages: [LlmMessage.user('Tell me a joke')],
  tools: const [],
)) {
  switch (event) {
    case LlmStreamText(:final delta): print(delta);
    case LlmStreamToolCall(:final call): print('Tool: ${call.name}');
    case LlmStreamDone(:final cachedTokenCount): print('Cached: $cachedTokenCount');
  }
}

// One-shot JSON output against a schema:
final result = await provider.generateStructured(
  prompt: 'Return a person record for "Marie Curie".',
  schema: Schema.object(properties: {
    'name': Schema.string(),
    'born': Schema.integer(),
    'fields': Schema.array(items: Schema.string()),
  }),
);
```

### Gemini (legacy, deprecated)

```dart
@Deprecated('Use FirebaseAiProvider — google_generative_ai is archived upstream.')
GeminiProvider(
  apiKey: 'your-gemini-api-key',
  model: 'gemini-2.0-flash',
  temperature: 0.2,
  requestTimeout: Duration(seconds: 45),
)
```

Uses the archived `google_generative_ai` package. Will be removed in v0.3.0. See [Migrating from `GeminiProvider`](#migrating-from-geminiprovider) below.

### Claude (Anthropic)

```dart
ClaudeProvider(
  apiKey: 'sk-ant-...',
  model: 'claude-sonnet-4-20250514',  // default
  maxTokens: 4096,
  temperature: 0.2,
  baseUrl: 'https://api.anthropic.com/v1',  // override for proxies
)
```

Uses the Anthropic Messages API via HTTP. No additional SDK dependency.

### OpenAI (GPT-4)

```dart
OpenAiProvider(
  apiKey: 'sk-...',
  model: 'gpt-4o',               // default
  temperature: 0.2,
  baseUrl: 'https://api.openai.com/v1',  // override for Azure
)
```

Uses the Chat Completions API via HTTP. Override `baseUrl` for Azure OpenAI or proxies.

### Bring Your Own Provider

Implement the `LlmProvider` interface:

```dart
abstract class LlmProvider {
  Future<LlmResponse> sendMessage({
    required List<LlmMessage> messages,
    required List<ToolDefinition> tools,
    String? systemPrompt,
  });

void dispose() {}
}
```

The package handles conversation format, tool schemas, and response parsing — your provider just needs to translate between `LlmMessage`/`LlmResponse` and your API's format.

**Built-in error handling**: All HTTP providers share typed exceptions (`RateLimitException`, `ContextOverflowException`, `AuthenticationException`, `ContentFilteredException`) and automatic retry with exponential backoff on rate limits.

---

## Migrating from `GeminiProvider`

`GeminiProvider` is deprecated as of v0.2.0 and will be removed in v0.3.0. Migrate to `FirebaseAiProvider`:

1. Add `firebase_core` and `firebase_ai` to your pubspec (`firebase_app_check` is optional but recommended).
2. Initialise Firebase before `runApp(...)`. Most apps already do this for analytics/auth/messaging.
3. Replace the provider construction:

```dart
   // BEFORE
   provider: GeminiProvider(apiKey: myGeminiKey),

// AFTER (no App Check — quickest migration, removes the in-binary key)
   provider: FirebaseAiProvider(
     firebaseAi: FirebaseAI.googleAI(),
     model: 'gemini-2.5-flash',
   ),

// AFTER (with App Check — recommended for production)
   provider: FirebaseAiProvider(
     firebaseAi: FirebaseAI.googleAI(
       appCheck: FirebaseAppCheck.instance,
       useLimitedUseAppCheckTokens: true,
     ),
     model: 'gemini-2.5-flash',
   ),
   ```

4. **Revoke** any Gemini API key that previously shipped in your app. Even after the migration, the old keys remain in distributed binaries and must be rotated in [Google AI Studio](https://aistudio.google.com/app/apikey).
5. Remove the API key from your `.env` files and any backend pipelines that fetched it. The Firebase project's stored Gemini API key (configured in the Firebase Console under "AI Logic") replaces it.

## Best Practices

### Securing LLM API Keys

**Never ship API keys in your app binary.** Hardcoded keys can be extracted from your APK/IPA in minutes. The recommended path is `FirebaseAiProvider` with App Check (see above), which keeps the Gemini API key entirely off the device.

For Claude or OpenAI (or any other HTTP-based provider), prefer **proxying through your backend** so no LLM key ships in the app:

```dart
// Your backend proxies calls to Claude/OpenAI
// This way the LLM key never leaves your server
ClaudeProvider(
  apiKey: userSessionToken,              // your own auth token
  baseUrl: 'https://api.yourapp.com/ai', // your proxy endpoint
)
```

This also lets you add rate limiting, cost tracking, content filtering, and audit logging server-side.

### Handling Sensitive Screens

If your app has screens with sensitive data (bank details, passwords) that the assistant shouldn't read or act on, you can exclude Semantics nodes using Flutter's built-in `ExcludeSemantics` widget:

```dart
ExcludeSemantics(
  child: CreditCardForm(...),
)
```

The assistant only sees what's in the Semantics tree — excluded subtrees are invisible to it.

### Domain Instructions

Invest time in `domainInstructions` and `fewShotExamples`. These are the highest-leverage configuration options — a few well-written examples dramatically improve the agent's accuracy and reduce unnecessary tool calls. Write them for your most common user flows first.

---

## Custom Tools

```dart
AiAssistantConfig(
  provider: FirebaseAiProvider(firebaseAi: FirebaseAI.googleAI()),
  customTools: [
    AiTool(
      name: 'check_inventory',
      description: 'Check if a product is in stock and get its current price.',
      parameters: {
        'productName': const ToolParameter(
          type: 'string',
          description: 'The name of the product to check.',
        ),
      },
      required: ['productName'],
      handler: (args) async {
        final name = args['productName'] as String;
        final result = await inventoryService.check(name);
        return {
          'inStock': result.inStock,
          'price': result.price,
          'quantity': result.available,
        };
      },
    ),
  ],
)
```

The LLM sees your tool's name, description, and parameters alongside the built-in tools. When it decides your tool is relevant, it calls it — the return map is fed back as the tool result.

### Built-in Tools

These are registered automatically and work on any Flutter app:

| Tool | What it does |
|------|-------------|
| `tap_element` | Taps a button, link, or interactive element by its label. Supports `parentContext` for disambiguation. |
| `set_text` | Enters text into any text field — auto-finds hidden/unfocused fields, activates search bars. |
| `scroll` | Scrolls up, down, left, or right to find off-screen content. |
| `navigate_to_route` | Navigates to a named route (e.g. `/store`, `/cart`). |
| `go_back` | Pops the current route (back button). |
| `get_screen_content` | Re-reads the current screen's semantics tree for fresh context. |
| `long_press_element` | Long presses an element (for context menus, etc.). |
| `increase_value` | Increases a quantity stepper or slider value. |
| `decrease_value` | Decreases a quantity stepper or slider value. |
| `ask_user` | Asks the user a question and waits for their response. Used only when genuinely ambiguous. |
| `hand_off_to_user` | Clears the overlay so the user can tap the final irreversible action button themselves. Only available when `confirmDestructiveActions: true` (default). |

---

## App Manifest (Code Generation)

For large apps, you can generate a rich "building map" that gives the assistant detailed knowledge of every screen without having to visit them first.

The recommended auth path uses Vertex AI + Application Default Credentials, so no long-lived Gemini key needs to live in `.env` or CI variables. Run `gcloud auth application-default login` once on the developer machine (or set `GOOGLE_APPLICATION_CREDENTIALS` to a service-account JSON in CI), then:

```bash
dart run flutter_ai_assistant:generate \
  --project=YOUR_GCP_PROJECT \
  --routes-file=lib/models/routes.dart \
  --router-file=lib/app/router.dart \
  --output=lib/ai_app_manifest.g.dart
```

Legacy auth (still supported, emits a deprecation warning):

```bash
dart run flutter_ai_assistant:generate \
  --api-key=YOUR_GEMINI_KEY \
  --routes-file=lib/models/routes.dart \
  --router-file=lib/app/router.dart \
  --output=lib/ai_app_manifest.g.dart

# Or with an env file containing GEMINI_API_KEY=...:
dart run flutter_ai_assistant:generate --env=.env.staging
```

This scans your route definitions and widget source code, sends each screen to Gemini for analysis, and generates a Dart file containing an `AiAppManifest` with:

- **Screen descriptions** — what each screen does, its sections, and interactive elements
- **Navigation links** — how screens connect to each other
- **Multi-step flows** — common user journeys spanning multiple screens
- **Global navigation** — bottom nav tabs, side menu structure

Pass the generated manifest to the config:

```dart
import 'ai_app_manifest.g.dart';

AiAssistantConfig(
  provider: FirebaseAiProvider(firebaseAi: FirebaseAI.googleAI()),
  appManifest: aiAppManifest,
)
```

The manifest provides a two-tier context system:
1. **Tier 1 (always loaded)**: App overview, all screens, navigation structure, flows
2. **Tier 2 (on-demand)**: Detailed screen sections, elements, and actions for the current screen

---

## Voice I/O

### Speech-to-Text (Input)

The assistant uses `speech_to_text` with intelligent locale resolution:

```dart
AiAssistantConfig(
  voiceEnabled: true,
  preferredLocales: ['en_US', 'hi_IN', 'es_ES'],  // your priority order
)
```

- On first listen, queries the device for available locales
- Picks the best match from your preferred list (supports exact match, hyphen variants, and language prefix matching)
- Partial transcription shown live as the user speaks
- Low-confidence results are filtered before reaching the LLM

### Text-to-Speech (Output)

The assistant uses `flutter_tts` with automatic language detection:

```dart
AiAssistantConfig(
  enableTts: true,
)
```

- **Auto-detects language**: Devanagari script or Hindi/Hinglish particles -> Hindi voice; otherwise English
- **Summary mode**: Long responses are truncated to the first sentence for TTS; full text stays in chat
- **Progress updates**: During multi-step tasks, the assistant speaks status updates ("Opening the store...", "Searching for onions...")

---

## Analytics Events

Every significant action in the assistant lifecycle emits a structured `AiEvent`:

```dart
AiAssistantConfig(
  onEvent: (event) {
    analytics.logEvent(
      name: 'ai_${event.type.name}',
      parameters: event.properties.map(
        (k, v) => MapEntry(k, v?.toString() ?? ''),
      ),
    );
  },
)
```

### Event Categories

**Conversation lifecycle**: `conversationStarted`, `conversationCompleted`, `conversationError`, `messageSent`, `messageReceived`, `conversationCleared`

**Agent loop**: `agentIterationStarted`, `agentIterationCompleted`, `agentCancelled`, `agentTimeout`, `agentMaxIterationsReached`, `agentOrientationCheckpoint`, `agentCircuitBreakerFired`

**LLM communication**: `llmRequestSent`, `llmResponseReceived`, `llmError`, `llmEmptyResponse`

**Tool execution**: `toolExecutionStarted`, `toolExecutionCompleted`, `screenContentCaptured`, `screenStabilizationAttempted`

**Voice**: `voiceInputStarted`, `voiceInputCompleted`, `voiceInputError`, `ttsStarted`

**UI interactions**: `chatOverlayOpened`, `chatOverlayClosed`, `suggestionChipTapped`, `buttonTapped`, `handoffStarted`, `handoffCompleted`, `askUserStarted`, `askUserCompleted`, `stopRequested`, `responsePopupShown`

**Navigation**: `routeChanged`, `navigationExecuted`

Each event includes relevant properties (documented on the `AiEventType` enum). For example, `toolExecutionCompleted` includes `toolName`, `arguments`, `success`, `error`, `durationMs`, and `iteration`.

---

## Rich Chat Content

Chat messages support rich content blocks beyond plain text:

### Interactive Buttons

```dart
// Post-task follow-up chips
AiAssistantConfig(
  postTaskChipsBuilder: (response) {
    final addedToCart = response.actions.any((a) =>
      a.toolName == 'tap_element' &&
      (a.arguments['label'] as String?)?.contains('ADD') == true);

if (addedToCart) {
      return ButtonsContent(
        buttons: [
          ChatButton(label: 'View my cart', icon: Icons.shopping_cart),
          ChatButton(label: 'Continue shopping', icon: Icons.store),
        ],
      );
    }
    return null;
  },
)
```

### Suggestion Chips

```dart
AiAssistantConfig(
  initialSuggestions: [
    AiSuggestionChip(
      icon: Icons.directions_car,
      label: 'Book a ride',
      message: 'Book me a ride',
    ),
    AiSuggestionChip(
      icon: Icons.storefront,
      label: 'Browse store',
      message: 'Take me to the store',
    ),
    AiSuggestionChip(
      icon: Icons.account_balance_wallet,
      label: 'Check balance',
      message: 'Show my wallet balance',
    ),
  ],
)
```

### Content Types

| Type | Description |
|------|-------------|
| `TextContent` | Plain text block |
| `ImageContent` | Inline image (URL or bytes) with optional caption |
| `ButtonsContent` | Group of tappable quick-reply buttons (wrap or column layout) |
| `CardContent` | Rich card with title, subtitle, image, and action buttons |

Button styles: `primary` (filled), `outlined` (default), `destructive` (red), `success` (green).

---

## Architecture

```
flutter_ai_assistant/
  lib/
    flutter_ai_assistant.dart    # Barrel file — public API
    src/
      core/
        ai_assistant.dart        # AiAssistant widget (wrap your app)
        ai_assistant_config.dart # All configuration options
        ai_assistant_controller.dart # Central orchestrator
        ai_event.dart            # Analytics event system
        ai_logger.dart           # Debug logging
      llm/
        llm_provider.dart        # Abstract provider interface
        react_agent.dart         # ReAct agent loop
        conversation_memory.dart # Multi-turn memory management
        providers/
          gemini_provider.dart   # Google Gemini
          claude_provider.dart   # Anthropic Claude
          openai_provider.dart   # OpenAI GPT-4
      action/
        action_executor.dart     # Executes UI actions (tap, type, scroll)
        scroll_handler.dart      # Smart scrolling
      context/
        semantics_walker.dart    # Reads the Flutter Semantics tree
        screen_context.dart      # Structured screen representation
        context_cache.dart       # TTL-based caching
        context_invalidator.dart # Cache invalidation on UI changes
        route_discovery.dart     # Progressive route learning
        screenshot_capture.dart  # Screen capture for multimodal
        ai_navigator_observer.dart # NavigatorObserver for route tracking
      tools/
        built_in_tools.dart      # Built-in UI interaction tools
        tool_definition.dart     # Tool schema (AiTool, ToolParameter)
        tool_registry.dart       # Tool registration and lookup
        tool_result.dart         # Structured tool results
      manifest/
        ai_app_manifest.dart     # Hierarchical app description
        ai_screen_manifest.dart  # Per-screen detail
        ai_section_manifest.dart # Screen sections
        ai_element_manifest.dart # Interactive elements
        ai_flow_manifest.dart    # Multi-step user journeys
        ai_nav_entry.dart        # Global navigation entries
        ai_action_manifest.dart  # Screen-level actions
        ai_navigation_link.dart  # Screen-to-screen links
        manifest.dart            # Barrel file re-exporting all manifest types
      voice/
        voice_input_service.dart # Speech-to-text
        voice_output_service.dart # Text-to-speech with language detection
      ui/
        chat_overlay.dart        # Full chat UI
        chat_bubble.dart         # Message bubbles with rich content
        action_feed_overlay.dart # Live action progress feed
        handoff_indicator.dart   # Handoff mode indicator
        response_popup.dart      # Compact result popup above FAB
      models/
        chat_message.dart        # Chat message model
        chat_content.dart        # Rich content types (text, image, buttons, cards)
        action_step.dart         # Action progress tracking
        agent_action.dart        # Executed action records
        app_context_snapshot.dart # Full app state for LLM
        ui_element.dart          # UI element representation
  bin/
    generate.dart                # CLI manifest generator
```

### Data Flow

1. **User input** (text or voice) -> `AiAssistantController`
2. **Context capture** -> `SemanticsWalker` reads the UI tree, `ContextCache` manages freshness
3. **Agent loop** -> `ReactAgent` sends context + history + tools to the `LlmProvider`
4. **Tool execution** -> LLM returns tool calls, `ActionExecutor` performs them on the live UI
5. **Observation** -> Screen is re-read, results fed back to the LLM
6. **Repeat** until the LLM returns a text response (task complete)
7. **Verification** -> System re-checks the screen to confirm task completion
8. **Response** -> Shown in chat, spoken via TTS if voice-initiated

---

## API Reference

### Accessing the Controller

From anywhere in the widget tree below `AiAssistant`:

```dart
// With rebuild on changes (in build methods)
final ctrl = AiAssistant.of(context);

// Without rebuild (in callbacks, initState)
final ctrl = AiAssistant.read(context);
```

### Controller Methods

| Method | Description |
|--------|-------------|
| `sendMessage(String text, {bool isVoice})` | Send a text command to the assistant |
| `startVoiceInput()` | Activate the microphone |
| `stopVoiceInput()` | Stop listening |
| `toggleOverlay()` | Show/hide the chat overlay |
| `showOverlay()` | Show the chat overlay |
| `hideOverlay()` | Hide the chat overlay |
| `requestStop()` | Stop the current agent execution |
| `cancelHandoff()` | Cancel a pending handoff (user decides not to act) |
| `clearConversation()` | Clear all chat messages and memory |
| `dismissResponsePopup()` | Programmatically dismiss the response popup above the FAB |

### Controller State (Getters)

| Getter | Type | Description |
|--------|------|-------------|
| `messages` | `List<AiChatMessage>` | All chat messages |
| `isProcessing` | `bool` | Whether the agent is currently executing |
| `isListening` | `bool` | Whether the mic is active |
| `isOverlayVisible` | `bool` | Whether the chat overlay is showing |
| `isHandoffMode` | `bool` | Whether waiting for user to tap final action |
| `isWaitingForUserResponse` | `bool` | Whether an `ask_user` question is pending |
| `isActionFeedVisible` | `bool` | Whether the action feed is showing in the overlay |
| `isResponsePopupVisible` | `bool` | Whether the response popup is showing above the FAB |
| `progressText` | `String?` | Current agent status text |
| `finalResponseText` | `String?` | The final response text (for action feed display) |
| `actionSteps` | `List<ActionStep>` | Live action progress steps |
| `hasUnreadResponse` | `bool` | Whether there's an unread response (FAB badge) |
| `partialTranscription` | `String?` | Live partial speech recognition text while user speaks |
| `config` | `AiAssistantConfig` | The current configuration (read-only access) |
| `handoffButtonLabel` | `String?` | Label of the button the user should tap during handoff |
| `handoffSummary` | `String?` | Description of what happens when the user taps the handoff button |
| `responsePopupType` | `AiResponseType` | Type of the response popup (action confirmation vs info card) |
| `responsePopupText` | `String?` | Text content shown in the response popup |

### Controller Properties

| Property | Type | Description |
|----------|------|-------------|
| `navigatorObserver` | `AiNavigatorObserver` | Add this to your `MaterialApp.navigatorObservers` for route tracking |

---

## Platform Support

The package itself supports all six Flutter platforms. **Provider availability** varies because `firebase_ai` does not ship plugin bindings for Linux or Windows desktop:

| Platform | `FirebaseAiProvider` | `ClaudeProvider` / `OpenAiProvider` (HTTP) |
|----------|----------------------|--------------------------------------------|
| Android  | Yes                  | Yes |
| iOS      | Yes                  | Yes |
| Web      | Yes                  | Yes |
| macOS    | Yes                  | Yes |
| Linux    | **No** (use Claude/OpenAI) | Yes |
| Windows  | **No** (use Claude/OpenAI) | Yes |

If you target Linux or Windows desktop and need Gemini specifically, either pin to `flutter_ai_assistant: ^0.1.x` (uses the deprecated `google_generative_ai` package, which builds on those platforms) or proxy Gemini calls through your own backend and use `ClaudeProvider`-style HTTP transport.

Voice features (speech-to-text, text-to-speech) depend on platform availability. The assistant gracefully degrades to text-only on platforms without voice support.

---

## Requirements

- Flutter >= 3.32.0
- Dart SDK >= 3.8.1

## Dependencies

| Package | Purpose |
|---------|---------|
| `firebase_core` | Firebase initialisation (required by `firebase_ai`) |
| `firebase_ai` | Recommended Gemini provider (`FirebaseAiProvider`) — App Check, implicit caching |
| `google_generative_ai` | Legacy Gemini provider (`GeminiProvider`, deprecated, removed in v0.3.0) |
| `http` | Claude and OpenAI providers (HTTP API calls) |
| `speech_to_text` | Voice input |
| `flutter_tts` | Voice output |
| `uuid` | Unique message and action IDs |

---

## License

MIT — see [LICENSE](LICENSE) for details.

flutter_ai_assistant

flutter_ai_assistant 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

flutter_ai_assistant

flutter_ai_assistant 최신버젼 다운로드

오늘의 추천 프로젝트

Was this content helpful?

Recent Feedback

Download README