android-gemini-nano

Android Gemini Nano On-Device Integration 🧠

Gemini Nano allows for powerful, completely on-device AI generation. Because it runs locally via the AICore Android system service, careful life-cycle and asynchronous resource management is required.

⚡ When to Use

• When building chat interfaces, generating summaries, or drafting text without cloud latency.
• When the user asks to integrate "Gemini Nano" or "AICore".

🏗️ Core Rules / Pillars

1. AICore Dependency and Availability Check

• Pattern: AICore might not be downloaded or fully initialized. Always check availability before starting a session.
• Implementation:

  val generativeModel = GenerativeModel("gemini-nano")
  
  // ALWAYS verify availability first
  val isAvailable = generativeModel.isAvailable()
  if (!isAvailable) {
      // Request download or show fallback UI
  }
  

2. Streaming Responses with Flow

• Pattern: Because generation can take time on low-end devices, always use generateContentStream to stream the UI update instead of waiting for the full response.
• Anti-Pattern: NEVER block the main thread.
• Implementation:

  viewModelScope.launch {
      generativeModel.generateContentStream(prompt)
          .catch { e -> handleError(e) }
          .collect { chunk ->
              _uiState.update { it.copy(text = it.text + chunk.text) }
          }
  }
  

🚧 Critical Anti-Hallucination Guards

• Trap: GenerativeModel is part of the com.google.ai.client.generativeai SDK. Do NOT hallucinate custom AICoreManager wrapper singletons.

🔗 Related Resources

• Google AI Edge / AICore SDK