GenAI Fireside Chat Application (2025)

Overview

A multi-AI chat application showcasing integration with 5 different AI providers:

ChatGPT (GPT-4.1 / GPT-4.1-mini) - OpenAI's latest models
Gemini 2.0 Flash - Google's Gemini via Firebase AI Logic
DeepSeek - AI model via Firebase Genkit & GitHub Models
Llama 3.1 8B - Meta's open-source model via Firebase Genkit
Gemma 3 Nano - On-device AI with multimodal support

This is a Flutter workshop project demonstrating modern AI integration patterns, streaming responses, and hybrid cloud/on-device architectures.

Prerequisites

Flutter SDK: 3.27+ (installation guide)
Dart SDK: 3.8+ (included with Flutter)
Node.js: 20+ for Firebase Cloud Functions (download)
IDE: Android Studio, VS Code, or IntelliJ with Flutter plugins
Firebase Account: For Gemini, DeepSeek, and Llama services
API Keys: OpenAI account for ChatGPT

Quick Start

1. Clone and Install Dependencies

# Clone the repository
git clone <repository-url>
cd ai_fireside_chat

# Install Flutter dependencies
flutter pub get

# Install Firebase Functions dependencies
cd genkit
npm install
cd ..

2. Configure API Keys

Create config/config.json from the template:

cp config/config.json.template config/config.json

Edit config/config.json and add your API keys:

{
    "chatGptApiKey": "sk-proj-...",
    "geminiApiKey": "",
    "claudeApiKey": ""
}

Note: geminiApiKey and claudeApiKey are not currently used. Gemini uses Firebase AI Logic (no API key needed), and Claude is not implemented.

3. Run the Application

# Run with configuration
flutter run --dart-define-from-file=config/config.json

# Or use your IDE's run button with configuration:
# VS Code: Set args in launch.json
# Android Studio: Edit Run Configuration

AI Provider Setup

🤖 ChatGPT (GPT-4.1)

Status: ✅ Ready to use

Setup:

Get an API key from OpenAI Platform
Add the key to config/config.json as chatGptApiKey
The app uses two API modes:
- Chat Completions API (gpt-4.1) - Direct streaming chat
- Assistants API (gpt-4.1-mini) - Persistent conversations

Implementation: lib/service/chat_gpt_service.dart and chat_gpt_completions_service.dart

Models:

gpt-4.1 - Most capable, 1M token context
gpt-4.1-mini - Cost-effective, 50% cheaper

🔮 Gemini 2.0 Flash (Firebase AI Logic)

Status: ✅ Ready to use

What is Firebase AI Logic? Firebase AI Logic (formerly Vertex AI in Firebase) provides seamless access to Google's Gemini models with automatic authentication, no API key management, and hybrid on-device/cloud inference.

Setup:

Connect to Firebase:

# Install Firebase CLI
npm install -g firebase-tools

# Login to Firebase
firebase login

# Initialize Firebase in your project
firebase init

Add Firebase to Flutter:
- Follow the FlutterFire setup guide
- Download google-services.json (Android) and GoogleService-Info.plist (iOS)
- Place them in the appropriate platform directories

Dependencies are already included:

firebase_core: ^4.2.0
firebase_ai: ^3.4.0

Firebase is already initialized in lib/main.dart

Implementation: lib/service/gemini_service.dart

Features:

✅ Automatic authentication (no API key needed)
✅ Streaming responses with generateContentStream()
✅ System instructions support
✅ Hybrid inference (on-device when available, cloud fallback)
✅ Model: gemini-2.0-flash (latest, fastest)

Benefits:

No API key management
Bundled Firebase pricing
Automatic security with App Check
Generous free tier

🧠 DeepSeek & Llama 3.1 (Firebase Genkit)

Status: ✅ Ready to use (requires Firebase setup)

What is Firebase Genkit? Firebase Genkit is an AI orchestration framework that allows you to build, deploy, and manage AI flows in Firebase Cloud Functions. It supports multiple AI providers with unified APIs.

Setup:

1. Initialize Firebase Functions

cd genkit

# Install dependencies (already done if you followed Quick Start)
npm install

# Login to Firebase (if not already)
firebase login

# Initialize Firebase project
firebase use --add
# Select your Firebase project or create a new one

2. Set Up GitHub Token Secret

Both DeepSeek and Llama use GitHub Models API through Genkit:

# Set GitHub token as a Firebase secret
firebase functions:secrets:set GITHUB_TOKEN

# Paste your GitHub Personal Access Token when prompted
# Get token from: https://github.com/settings/tokens
# Required scope: public_repo (or repo for private repos)

3. Deploy Cloud Functions

# Build TypeScript
npm run build

# Deploy to Firebase
npm run deploy

# Or use Firebase CLI directly
firebase deploy --only functions

4. Test Locally (Optional)

# Start Genkit development server
npm run genkit:start

# Or start Firebase emulators
npm run serve

Implementation:

Backend: genkit/src/chat.ts, genkit/src/flows.ts, genkit/src/index.ts
Frontend: lib/service/genkit_service.dart

Architecture:

Flutter App → Cloud Functions (Genkit) → GitHub Models API
                                       → Google AI (streaming fallback)

Models:

DeepSeek: Uses mistralSmall (non-streaming) or gemini-1.5-flash (streaming)
Llama 3.1 8B: Uses Meta-Llama-3.1-8B-Instruct or gemini-1.5-pro (streaming)

Features:

✅ Dual AI system (GitHub Models + Google AI)
✅ Session management
✅ Streaming support
✅ Automatic model selection based on streaming mode

Cloud Functions Endpoints:

initChatSession - Create new chat session
chat - Send message and get response
chatStream - Send message with streaming response

💎 Gemma 3 Nano (On-Device AI)

Status: ✅ Ready to use (requires model download)

What is Gemma? Gemma is Google's family of lightweight, open-source AI models designed to run efficiently on mobile devices. Gemma 3 Nano includes multimodal support (text + images).

Setup:

1. Dependencies are already included:

flutter_gemma: ^0.11.5

2. Download Model Files

Gemma 3 Nano models need to be downloaded separately due to their size (1-4GB).

Option A: Download Pre-converted Models

# Visit flutter_gemma releases
# https://github.com/DenisovAV/flutter_gemma/releases

# Download one of:
# - gemma-3-nano-1b-it.task (~1GB)
# - gemma-3-nano-2b-it.task (~2GB) - Recommended
# - gemma-3-nano-4b-vision.task (~4GB) - Multimodal support

# Place in assets directory
mkdir -p assets
mv gemma-3-nano-2b-it.task assets/

Option B: Convert Models Yourself

# Follow the guide at:
# https://pub.dev/packages/flutter_gemma#model-conversion

# Use MediaPipe Model Maker to convert Gemma models to .task format

3. Update pubspec.yaml

Assets are already configured:

flutter:
  assets:
     - assets/

4. Model Configuration

The app is configured to use:

Model file: gemma3-1b-it-int4.task (1B quantized)
Location: assets/gemma3-1b-it-int4.task

To use a different model, update lib/service/gemma_service.dart:

_gemma = await GemmaModel.load(
  modelPath: 'assets/your-model-name.task',
  modelType: GemmaModelType.gemma3Nano2bInstruct, // Adjust based on model
);

Implementation: lib/service/gemma_service.dart

Supported Models (flutter_gemma 0.11.5):

✅ Gemma 3 Nano (1B, 2B, 4B)
✅ Gemma 3 270M (ultra-lightweight)
✅ Llama 3.2 1B
✅ Phi-4 (Microsoft)
✅ DeepSeek, Qwen2.5, TinyLlama

Model Formats:

.task - MediaPipe-optimized (recommended)
.litertlm - Web-optimized
.bin / .tflite - Standard format

Features:

✅ Runs entirely on device (no internet required)
✅ Privacy-first (no data sent to servers)
✅ Zero cost per inference
✅ Low latency (~200ms first token on iPhone 15 Pro)
✅ GPU acceleration support
⚠️ Requires ~2GB RAM during inference

Performance (Gemma 3 Nano 2B on iPhone 15 Pro):

First token: ~200ms
Throughput: ~15 tokens/sec
Memory: ~2GB RAM
Battery: ~5% per 1000 tokens

Project Structure

ai_fireside_chat/
├── lib/
│   ├── bloc/               # State management (BLoC pattern)
│   │   └── chat_bloc.dart
│   ├── core/               # Domain models
│   │   ├── chat_message.dart
│   │   └── message_producer.dart
│   ├── service/            # AI service implementations
│   │   ├── chat_service.dart          # Abstract base class
│   │   ├── chat_gpt_service.dart      # GPT-4.1 (Assistants API)
│   │   ├── chat_gpt_completions_service.dart  # GPT-4.1 (Chat API)
│   │   ├── gemini_service.dart        # Gemini 2.0 Flash (firebase_ai)
│   │   ├── gemma_service.dart         # Gemma 3 Nano (on-device)
│   │   └── genkit_service.dart        # DeepSeek & Llama (Cloud Functions)
│   ├── widgets/            # UI components
│   └── main.dart           # App entry point
│
├── genkit/                 # Firebase Genkit backend
│   ├── src/
│   │   ├── chat.ts         # Session management & model orchestration
│   │   ├── flows.ts        # Genkit flow definitions
│   │   └── index.ts        # Cloud Functions exports
│   ├── package.json        # Node.js dependencies
│   └── tsconfig.json       # TypeScript configuration
│
├── config/
│   ├── config.json.template  # API keys template
│   └── config.json           # Your API keys (gitignored)
│
├── assets/                 # Model files (download separately)
│   └── *.task              # Gemma models (gitignored)
│
├── pubspec.yaml            # Flutter dependencies
└── README.md               # This file

Architecture Overview

State Management: BLoC Pattern

User Input → ChatBloc → ChatService → AI Provider → Stream Response → UI

ChatBloc: Manages conversation state and orchestrates service calls
ChatService: Abstract interface for all AI services
Concrete Services: ChatGPT, Gemini, DeepSeek, Llama, Gemma implementations

AI Service Abstraction

All AI services implement a common interface:

abstract class ChatService {
  Future<void> init({required String systemInstructions});
  Stream<String> processMessageStream(List<ChatMessage> messages);
  void dispose();
}

This allows swapping providers without changing UI code.

Streaming Architecture

All services support streaming responses for better UX:

await for (final chunk in service.processMessageStream(messages)) {
  // Update UI with each chunk as it arrives
}

Hybrid Cloud/Device Strategy

Cloud Services (ChatGPT, Gemini, DeepSeek, Llama):
- Best for complex reasoning
- Always available
- Higher cost per inference
- Requires internet
On-Device (Gemma):
- Best for privacy
- Zero cost
- Works offline
- Limited by device capabilities

Package Versions (2025)

Flutter Dependencies

dependencies:
  flutter_sdk: flutter

  # State Management
  flutter_bloc: ^8.0.1
  equatable: ^2.0.3

  # Firebase & AI
  firebase_core: ^4.2.0          # Firebase initialization
  firebase_ai: ^3.4.0            # Gemini 2.0 Flash (Firebase AI Logic)
  cloud_functions: ^6.0.3        # Call Genkit Cloud Functions

  # On-Device AI
  flutter_gemma: ^0.11.5         # Gemma 3 Nano with multimodal support

  # Utilities
  http: ^1.5.0                   # HTTP client for ChatGPT
  flutter_markdown: ^0.7.6+2     # Markdown rendering
  cupertino_icons: ^1.0.6        # iOS icons
  url_launcher: ^6.0.6           # Open URLs

dev_dependencies:
  flutter_test: flutter
  flutter_lints: ^6.0.0          # Linting rules

Node.js Dependencies (Genkit)

{
  "dependencies": {
    "genkit": "^1.21.0",                    // Core framework
    "@genkit-ai/firebase": "^1.21.0",       // Firebase integration
    "@genkit-ai/googleai": "^1.21.0",       // Google AI models
    "@genkit-ai/compat-oai": "^1.21.0",     // OpenAI compatibility
    "genkitx-github": "^1.15.0",            // GitHub Models plugin
    "firebase-admin": "^13.5.0",            // Admin SDK
    "firebase-functions": "^6.5.0",         // Cloud Functions Gen2
    "express": "^4.21.2",                   // Web framework
    "typescript": "^5.9.3"                  // TypeScript compiler
  }
}

Development Workflow

Local Development

# Run Flutter app
flutter run --dart-define-from-file=config/config.json

# Run with hot reload
flutter run --dart-define-from-file=config/config.json --hot

# Run on specific device
flutter devices
flutter run -d <device-id> --dart-define-from-file=config/config.json

Testing Cloud Functions Locally

cd genkit

# Start Firebase emulators
npm run serve

# Or start Genkit dev UI
npm run genkit:start

Debugging

# Flutter analyze
flutter analyze

# Check for outdated packages
flutter pub outdated

# Clean build
flutter clean
flutter pub get

Building for Production

# Android
flutter build apk --dart-define-from-file=config/config.json
flutter build appbundle --dart-define-from-file=config/config.json

# iOS
flutter build ios --dart-define-from-file=config/config.json

# macOS
flutter build macos --dart-define-from-file=config/config.json

Troubleshooting

ChatGPT: "Invalid API Key"

Verify your API key in config/config.json
Ensure you're using the --dart-define-from-file flag
Check that your OpenAI account has credits

Gemini: "Firebase not initialized"

Run firebase init in project root
Add google-services.json (Android) and GoogleService-Info.plist (iOS)
Verify firebase_core is initialized in main.dart

Genkit: "Function not found"

Deploy Cloud Functions: cd genkit && npm run deploy
Check Firebase project is selected: firebase use --add
Verify secrets are set: firebase functions:secrets:access GITHUB_TOKEN

Gemma: "Model not found"

Download model files from flutter_gemma releases
Place .task files in assets/ directory
Run flutter pub get to update asset manifest
Check model path in gemma_service.dart matches filename

Streaming not working

Check internet connection (for cloud services)
Verify processMessageStream() is used (not single-shot API)
For Genkit: ensure streaming is enabled in flow configuration

Cost Considerations

Provider	Model	Cost per 1M tokens	Best For
On-Device	Gemma 3 Nano	$0	Privacy, offline, high-volume
Firebase AI	Gemini 2.0 Flash	$	General chat, multimodal
OpenAI	GPT-4.1-nano	$	Fast responses
OpenAI	GPT-4.1-mini	$$	Balanced quality/cost
Genkit	Llama 3.1 8B	$ (via GitHub)	Open source
Genkit	DeepSeek	$$	Specialized tasks
OpenAI	GPT-4.1	$$$	Complex reasoning

Recommendations:

Use Gemma for < 512 tokens, offline scenarios, or privacy-critical data
Use Gemini 2.0 Flash for general conversations (best value)
Use GPT-4.1-mini for tasks requiring OpenAI capabilities
Use Llama/DeepSeek for open-source requirements

Learn More

Flutter Resources

Firebase & AI Resources

Workshop Resources

MODERNIZATION_PLAN_2025.md - Migration details
Architecture Analysis Report - Technical deep dive

Contributing

This is a workshop project. Feel free to:

Add new AI providers
Improve error handling
Add unit tests
Enhance UI/UX
Optimize performance

License

[Specify your license here]

Acknowledgments

Flutter Team - Amazing framework
Firebase Team - Seamless backend integration
OpenAI, Google, Meta - Powerful AI models
flutter_gemma Contributors - On-device AI plugin

Generated with Claude Code 🤖 Last Updated: October 2025

Name		Name	Last commit message	Last commit date
Latest commit History 21 Commits
.claude		.claude
.vscode		.vscode
android		android
assets		assets
config		config
genkit		genkit
ios		ios
lib		lib
linux		linux
macos		macos
web		web
windows		windows
.firebaserc		.firebaserc
.gitignore		.gitignore
.metadata		.metadata
CLAUDE.md		CLAUDE.md
MODERNIZATION_PLAN_2025.md		MODERNIZATION_PLAN_2025.md
README.md		README.md
analysis_options.yaml		analysis_options.yaml
firebase.json		firebase.json
pubspec.lock		pubspec.lock
pubspec.yaml		pubspec.yaml

DenisovAV/ai_fireside_chat

Folders and files

Latest commit

History

Repository files navigation

GenAI Fireside Chat Application (2025)

Overview

Prerequisites

Quick Start

1. Clone and Install Dependencies

2. Configure API Keys

3. Run the Application

AI Provider Setup

🤖 ChatGPT (GPT-4.1)

🔮 Gemini 2.0 Flash (Firebase AI Logic)

🧠 DeepSeek & Llama 3.1 (Firebase Genkit)

1. Initialize Firebase Functions

2. Set Up GitHub Token Secret

3. Deploy Cloud Functions

4. Test Locally (Optional)

💎 Gemma 3 Nano (On-Device AI)

1. Dependencies are already included:

2. Download Model Files

3. Update pubspec.yaml

4. Model Configuration

Project Structure

Architecture Overview

State Management: BLoC Pattern

AI Service Abstraction

Streaming Architecture

Hybrid Cloud/Device Strategy

Package Versions (2025)

Flutter Dependencies

Node.js Dependencies (Genkit)

Development Workflow

Local Development

Testing Cloud Functions Locally

Debugging

Building for Production

Troubleshooting

ChatGPT: "Invalid API Key"

Gemini: "Firebase not initialized"

Genkit: "Function not found"

Gemma: "Model not found"

Streaming not working

Cost Considerations

Learn More

Flutter Resources

Firebase & AI Resources

Workshop Resources

Contributing

License

Acknowledgments

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Languages

Packages