AI Mobile Chat App

This is a modern Android application that demonstrates on-device LLM inference using MediaPipe GenAI and Google AI Edge Local Agents. It features a conversational interface with support for native tool triggering (skills) like web search and calendar viewing.

The app uses Skill.md to declare AI Tools: view_calendar and web_search. SKILLs are placed in feature/chat/src/main/assets chat/util/SkillLoader.kt is used to find and parse Skills.

Getting Started for Developers

Follow these steps to set up the project on your local machine.

1. Install Android Studio

Download and install the latest stable version Android Studio Quail 1 of Android Studio (Ladybug or newer recommended).

2. Clone or Unzip the Project

Clone the repository from GitHub or unzip the provided archive into your preferred workspace directory.

3. Configure Local Properties

The project requires a Hugging Face User Access Token to download models.

1. Generate a token at huggingface.co/settings/tokens ( Read access is sufficient).
2. Open (or create) the local.properties file in the root directory of the project.
3. Add the following line:

   HUGGING_FACE_TOKEN=your_token_here

4. Open and Sync

1. Launch Android Studio.
2. Select Open and navigate to the project root folder.
3. Wait for the Gradle sync to complete.

Project Structure

The project follows a modular architecture:

Core Modules (:core)

• :core:coroutines: Base classes and utilities for Kotlin Coroutines and Flow-based UseCases.
• :core:database: Room database implementation for storing chat history.
• :core:di: Koin dependency injection configuration and extensions.
• :core:filestore: Manages file operations in the app's internal and cache directories.
• :core:navigation: Centralized navigation logic using Jetpack Compose Navigation.
• :core:ui: Shared UI components, themes (Material 3), and base ViewModel.
• :core:util: General-purpose utility classes and extensions.

Feature Modules (:feature)

• :feature:home: The main landing screen.
• :feature:chat: The conversational interface, including LLM inference logic and skill parsing.
• :feature:download: Background worker implementation for downloading LLM models from Hugging Face.
• :feature:settings: Application settings and configuration.
• :feature:tooling: Native tool implementations (e.g., Jsoup-based web search, Calendar provider).

App Module (:app)

• The main entry point that wires all modules together and initializes Koin.

Key Technologies

• Jetpack Compose: For a modern, declarative UI.
• MediaPipe GenAI: On-device LLM inference.
• Google AI Edge Local Agents: Structured function calling (skills) protocol.
• Koin: Lightweight dependency injection.
• WorkManager: Reliable background model downloading.
• Jsoup: Web scraping for the search tool.
• OkHttp: For network requests.

How to Run

1. Connect a physical Android device (recommended) or an emulator.
   • Note: LLM inference is resource-intensive and works best on modern devices (e.g., Pixel 7 or newer).
2. Select the app configuration in Android Studio.
3. Click Run.
4. Upon opening the Chat, the app will automatically start downloading the LLM model (~700MB). You can track the progress on the screen.

Issues:

1. A new Gemma 4 LLM doesn't work with MediaPipe engine. It requires LiteRT engine (more advanced implementation).
2. Sometimes LLM hallucinates when asking about agenda, view calendar instead of using ' view_calendar' tool

Author:

Dmitri Chernysh