I built an offline answer engine on top of kiwix-serve: every answer links to the ZIM page it came from
Hi all. Kiwix has been the backbone of a project I have been building for a few months, and it is now open source, so I wanted to show it here first.
The Ark of Noahledge node is off-grid AI for places where the grid is unreliable and the internet is not a given. It puts an open-weight language model beside a local library and answers questions from it. The part that depends on you: every claim in an answer carries a citation, and the citation opens the exact article in kiwix-serve on the same machine. What the model adds from its own memory is shown separately and labelled as unchecked.
What it does today:
- One command on a Windows PC with an NVIDIA card (6 GB or more) downloads a starter library (four ZIMs: zimgit medicine, post-disaster and water, and Wikipedia medicine in Spanish), checks every file by sha256, builds the index and answers a first question. 38 to 70 minutes in my tests.
- On my reference machine it serves 67 ZIMs from one library.xml, and searches 39 million passages drawn from 44 of them plus a few PDF collections.
Things I learned about kiwix-serve along the way (book keys, absolute paths in library.xml, the search form) are written up in the code comments; happy to compare notes or be told I did something the hard way.
Code (Apache 2.0): https://github.com/ArkOfNoahledge/ark-node More about the project: https://arkofnoahledge.org
Thank you for Kiwix. None of this would exist without it.