Ollama App

Android chat client for the Ollama Cloud API — streaming conversations, web search, image analysis, document processing, and LaTeX rendering.

Features

Chat & Streaming

• Real-time SSE streaming with token-by-token rendering
• Conversations persisted in Room with Flow-based reactive queries
• Auto-compact: transparent conversation summarization when approaching context limits
• Auto-generated titles (5 words max, toggleable)
• Foreground service keeps streaming alive when the screen is off

AI Tools

• web_search and web_fetch — the model decides when to search (up to 3 tool-calling rounds)
• Research agent executes searches and auto-fetches result pages
• Visual tool-call cards with green checkmarks and nested fetch results

Vision & Attachments

• Image analysis via compatible models (kimi-k2, gemma4, qwen3.5...)
• Vision fallback: non-vision models delegate to a configurable vision model via describe_image
• Attachments: images (gallery/camera/photo capture), PDFs, text files
• Smart image resize (max 1024px, JPEG quality 85%)

Rich Rendering

• Markdown: tables, headings, links, code blocks via mikepenz renderer
• LaTeX math: display ($$, \[) and inline ($, <latex>) via bundled KaTeX (0.16.21)
• Mixed LaTeX+Markdown rendered in a WebView for pixel-perfect output
• Text selection on all message bubbles

Model Management

• Search, enable/disable, custom context size per model
• Research agent model and vision fallback model configurable
• Recent models (10 most used)
• 25+ pre-configured models with default context sizes and vision/thinking capabilities

Internationalization

• Default language: English
• Full French UI translation (values-fr/)
• Language selector: System / English / Français / Español (no UI) / Deutsch (no UI)
• Adaptive system prompt injects the selected language, but always follows the user's question

Theme & UX

• Dark / Light / System theme
• Thinking mode toggle (shows reasoning content from deepseek models)
• Chat search with real-time filtering
• Drawer-based navigation with persistent ViewModels

Security

• API key: stored in EncryptedSharedPreferences (AES256-GCM, Android Keystore TEE/StrongBox)
• Cloud backup disabled (allowBackup=false) to prevent key leakage
• Prompt sanitization: 13 injection patterns blocked ([END OF SYSTEM PROMPT], <|im_start|>, etc.)
• URL validation (anti-SSRF): blocks localhost, private IPs (RFC 1918), link-local, local IPv6
• Conditional logging: debug/warning only in debug builds

Screenshots

[Add screenshots here]

Requirements

• Android 9.0 (API 28) or higher
• An Ollama API key

Installation

Download the latest APK from the releases page.

Build

export JAVA_HOME="/Applications/Android Studio.app/Contents/jbr/Contents/Home"
export ANDROID_HOME="$HOME/Library/Android/sdk"

./gradlew assembleDebug

The APK is generated at app/build/outputs/apk/debug/app-debug.apk.

# Install
adb install -r app/build/outputs/apk/debug/app-debug.apk

# Launch
adb shell am start -n org.canalweb.ollamaapp/.MainActivity

Tests

# Unit tests (35 tests)
./gradlew testDebugUnitTest

# Instrumentation tests (6 tests, requires emulator)
./gradlew connectedAndroidTest

Architecture

app/src/main/java/org/canalweb/ollamaapp/
├── MainActivity.kt
├── navigation/
│   └── AppNavigation.kt              # Compose NavHost (type-safe routes)
├── data/
│   ├── agent/
│   │   └── ResearchAgent.kt          # web_search + web_fetch agent
│   ├── api/
│   │   ├── OllamaApiService.kt       # Retrofit interface (chat, models)
│   │   ├── OllamaApiProvider.kt      # OkHttp/Retrofit singleton
│   │   ├── OllamaToolApiProvider.kt  # Tool API (web_search, web_fetch)
│   │   ├── RetryUtil.kt              # Fibonacci backoff retry (max 9, 1-55s)
│   │   └── SseStreamParser.kt        # Shared SSE parser
│   ├── database/                     # Room v7 (chats, messages, models)
│   ├── storage/
│   │   └── ChatStorage.kt            # Room facade
│   └── store/
│       └── SettingsDataStore.kt      # EncryptedSharedPreferences + DataStore
├── ui/
│   ├── theme/                        # Material3, dynamic colors
│   ├── screens/                      # ChatList, Chat, ModelSelection, ModelManagement, Settings, Welcome
│   └── viewmodels/                   # ChatList, Chat, ModelSelection, ModelManagement, Settings
└── utils/
    ├── AppLanguage.kt                # Language resolution
    ├── FileUtils.kt                  # Base64, image resize, PDF/text extraction
    ├── HtmlRenderer.kt               # CommonMark → HTML + KaTeX
    ├── Logger.kt                     # Conditional logging (DEBUG only)
    ├── TokenCounter.kt               # JTokkit-based token counting
    └── UrlValidator.kt               # Anti-SSRF (blocks private IPs)

Tech Stack

Category	Technology
Language	Kotlin
UI	Jetpack Compose + Material3
Database	Room (v7 with migrations)
Secure storage	EncryptedSharedPreferences (AES256-GCM)
HTTP	Retrofit + OkHttp
Markdown	multiplatform-markdown-renderer-m3 (mikepenz)
Markdown→HTML	CommonMark + commonmark-tables
LaTeX	KaTeX 0.16.21 (bundled JS/CSS assets)
Images	Coil Compose
PDF	pdfbox-android 2.0.27.0
Token counting	JTokkit
Build	Gradle 9.4.1 + KSP 2.2.10-2.0.2
Min SDK	API 28 (Android 9.0)
Target SDK	API 36

Supported Models

The app auto-detects available models via the Ollama API. Context sizes and capabilities (vision, thinking) are pre-configured for 25+ models including deepseek-v4, kimi-k2, gemma4, qwen3, glm, minimax, ministral, nemotron, gemini, and gpt-oss.

Acknowledgments

This app was developed with significant assistance from DeepSeek-v4-pro, an AI coding assistant that contributed to architecture decisions, code generation, debugging, and optimization throughout the project.

License

MIT