anonnerdfrenzy/survival-llm

Survival LLM is a desktop app for asking practical emergency and first-aid questions when the internet is unavailable. It pairs a locally run language model with a downloadable offline article library, and every answer links back to the passage it came from. It is aimed at people preparing for blackouts, outages or other situations where online help may not be reachable, not at developers integrating a model.

How a question is answered

A question goes to a local model served through Ollama. During the turn the model can search a SQLite full-text index built from roughly 1,500-character passages, open a full article, or show an article on screen; the ranking favors official sources. The chat page is served only on 127.0.0.1, so the exchange stays on the machine. The explainer diagram in the README shows this tool loop in detail How it works.

The distinguishing feature is traceability. Each answer carries a "Sources:" line, and clicking a source opens the article with the cited passage highlighted and its license. Because models have been observed naming sources they never opened, the app checks that line against the articles the model actually read or saw in search results and substitutes the real list when it differs citation test results. A hidden reasoning view can also be expanded, with a token budget controlling how much the model thinks before answering.

Setup and hardware

Installation has to happen while online; afterwards it works disconnected. The Mac app requires Apple Silicon and macOS 14 Sonoma or newer; Windows 10 or 11 64-bit is also supported. Setup downloads an AI model and a library, and the README estimates 20 minutes to about 2.5 hours depending on the machine and connection, with the largest model at 80 GB. Disk space needs the model plus 5 GB, plus peak space during library installation. The app chooses a model from available memory: 8 GB machines can search the library but are not offered a model, and the smallest models the authors tested were judged too unsafe to recommend install and model table.

Practical caveats

The library contains around 1.5 million articles, or 500,000 in the smaller download, and includes Wikipedia, medical references, government emergency guides, manuals and practical Q&A; content keeps its source license, and a default build including non-shareable material is for personal use only sources and licensing. Files the user adds are indexed separately, and memory notes about people are kept in plain text locally. The author's evaluation is modest: 22 blackout and first-aid questions graded by AI judges, run on an earlier library build holding some material not in the public downloads, so the reported table is a model comparison rather than a guarantee. The README states plainly that AI models make mistakes and that answers should be checked against the shown sources evaluation scope. Updates are fetched from the project's web page and verified before restarting into the new version.