← sources

Web search for a 1.7B model running fully on a phone, no search API and no backend (open source)

source: reddit (reddit:LocalLLaMA)published 2026-10-14 11:28ingested 2026-10-15 02:42

retained text

We just shipped web search in Private Mind, our open source app for running LLMs locally on iOS and Android. The video is the whole pitch: Qwen 3 1.7B with web off makes up Apple's September event, with web on it gets it right and cites 5 pages.

The fun constraint is that a 1.7B model with a short context can't just read 5 web pages. So most of the work is picking the few hundred words it actually gets to see.

Roughly how it works:

* There's no search API. A hidden WebView on the phone opens DuckDuckGo (Brave and Mojeek as fallbacks) and we parse the results page
* It opens the top pages and cuts them into \~500 character chunks
* An on-device embedding model (LFM2.5, \~415MB) plus some keyword matching picks the best 6 chunks, max 3 from one page
* If only one site claims something, it counts for less than something two sites agree on
* The model answers from those chunks and links every source

Web search is a per-chat toggle and off by default. When it's on, the phone makes the same requests a browser would. The chat itself stays on the device.

Everything runs on react-native-executorch, both the LLM and the embeddings.

Repo: [https://github.com/software-mansion-labs/private-mind](https://github.com/software-mansion-labs/private-mind)

document fetched 2026-10-15 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5oqiy/web_search_for_a_17b_model_running_fully_on_a