Private Mind adds on-device web search for Qwen 3 1.7B on iOS and Android

Private Mind, an open-source application for running LLMs locally on iOS and Android, has shipped a web search feature that operates entirely on-device. The implementation allows a Qwen 3 1.7B model to access real-time information, such as correcting hallucinations about recent events, while maintaining privacy by keeping chat data on the device.

How It Works

Since a 1.7B model has a limited context window, the system cannot ingest full web pages. Instead, it employs a multi-step retrieval process:

• No Search API: A hidden WebView on the phone opens DuckDuckGo (with Brave and Mojeek as fallbacks) to parse search results directly.
• Chunking: The top results are opened and split into approximately 500-character chunks.
• Selection: An on-device embedding model, LFM2.5 (~415MB), combined with keyword matching, selects the best 6 chunks (max 3 from any single page).
• Consensus: Information corroborated by multiple sites is weighted higher than single-source claims.
• Response: The model answers using only the selected chunks and links every source.

Technical Implementation

The entire pipeline, including both the LLM and the embedding model, runs on react-native-executorch. Web search is disabled by default and acts as a per-chat toggle. When enabled, the phone makes network requests identical to a standard browser, but the chat history and processing remain local.