You are not logged in.
Hi all,
I'd like to share Umbra Wiki, an offline assistant that answers questions from a local library using a local model. Nothing leaves the machine and it needs no account.
How it works
A small Python backend (standard library only) runs kiwix-serve over offline ZIM archives (Stack Exchange Q&As, WikEM, iFixit, Appropedia, energypedia, WikiMed and more) plus a built-in field manual.
It finds the most relevant passages and asks a local model through Ollama's API; every answer carries numbered citations you can open.
A GTK + WebKit (webkit2gtk-4.1) window shows the answers. Settings live in ~/.config/umbra-wiki, history in ~/.local/share/umbra-wiki.
The model (Gemma 3 1B/4B or Llama 3.1 8B) and the library packs (0.9 to 10.1 GB) are chosen in a first-run tour.
Packaging
Signed pacman repository hosted on GitHub Pages, with an install guide: https://umbraxc.github.io/umbra-repo/
Key fingerprint:
E50B 101F 9852 9A32 564F DD3B 37A6 D2DA 9E6A 0F0BThe PKGBUILD builds from a tagged, checksummed release and is namcap-clean. I'll submit it to the AUR as umbra-wiki once account registration reopens.
Depends on ollama, kiwix-tools, webkit2gtk-4.1, python-gobject; ollama-cuda or ollama-rocm are optional for GPUs.
On a CPU-only laptop (i5, Intel graphics) an answer takes about a minute with gemma3:4b.
Screenshots: start screen, an answer with sources, field manual
Source (MIT): https://github.com/umbraxc/omarchy-umbra
It was built with a lot of help from Claude (credited in the README). Feedback on the packaging or anything else is very welcome.
Offline
