On-device web search for a 1.7B phone-local LLM, no search API or backend
K4anan · reddit · 2026-10-14
The open-source app Private Mind added web search to fully local LLMs on iOS/Android. With web off, Qwen 3 1.7B hallucinates Apple's September event; with it on, it answers correctly and cites 5 pages. The hard constraint: a 1.7B model with a short context can't read 5 pages, so most of the work is choosing the few hundred words it actually sees.
- No search API: a hidden WebView opens DuckDuckGo (Brave/Mojeek fallbacks) and parses the results page.
- Top pages are split into 500-char chunks; an on-device embedding model (LFM2.5, 415MB) plus keyword matching picks the best 6 chunks, max 3 per page.
- Claims from a single site count less than claims two sites agree on; answers link every source.
- Web search is a per-chat toggle, off by default; the phone makes the same requests a browser would, and chat stays on device. Stack: react-native-executorch for both LLM and embeddings. Repo: software-mansion-labs/private-mind.
More from coding & agent
- Agent Baseline: Docker, Barclays et al. Publish Open Reference Architecture for Securing Autonomous Software Factories — AccBalanced · 2026-10-14
- He runs 5 one-person companies in Cursor toward $50K MRR with this OS — PrajwalTomar_ · 2026-10-14
- Meta paper: reward hacking spreads between agents in multi-agent coding systems — omarsar0 · 2026-10-14
- Read, prepare, execute: a three-tier framework for deciding what AI agents may do autonomously — Altadigit · 2026-10-14
- One Prompt, an Infinite 3D Musical Garden: 82,000 Grass Tufts and a Gesture-Played Harp — AI_Andrew · 2026-10-14
- Founder says AI teammates did 10x the work but created huge mental load with Squad — tibo_maker · 2026-10-14