Survival LLM: offline AI answer bot with 約93万件 Wikipedia library

Survival LLM is built around a simple trade-off: it gives up the reach of a cloud AI service in exchange for working when nothing else does. The setup step is the only online moment — the installer pulls down the AI engine, the model and the library, which takes 20 minutes to two and a half hours depending on the machine and connection speed, though the actual hands-on part is about five minutes.

The library is deliberately limited to openly licensed or public-domain material, which is what makes it distributable and instantly searchable. Alongside Wikipedia it bundles WikiMed, MedlinePlus, UK Medicines A to Z, WikEM emergency medicine, Ready.gov and CDC travel guidance, US military medical and survival manuals, wikiHow, iFixit repair guides, Appropedia and Stack Exchange Q&A on home repair, cooking, gardening, outdoors and amateur radio. Users can also load their own PDFs, Word documents, ebooks and notes into the search set, and those files stay local too.

The sample answers show the register the tool aims for. On refrigerator food during an outage it says a closed door keeps food safe about four hours, that a thermometer reading of 40°F(4℃) or below is still safe, and that anything above that for two hours should be thrown out — with an explicit warning never to taste food to check it. A full freezer holds about 48 hours, a half-full one about 24. On severe bleeding it walks through direct pressure, elevating the leg, adding gauze rather than peeling wet gauze away, watching for signs of shock, and a tourniquet with the time recorded as a last resort.

Power draw is modest: about 100W while generating an answer on a MacBook Pro, at 5 to 40 seconds per question, which the project says works out to roughly 100 to 300 questions on a full charge — enough to run from a small generator or portable power station. A built-in update button refreshes the app while keeping models, libraries and added files in place.