A library and an AI that work without the internet
Field: offline knowledge, local AI, open source software
Period: September 2026
Client: internal project, Lab 77 (R&D)
A complete encyclopedia, courses, books, and an assistant that answers your questions from those documents alone, all served by a machine sitting in the room, with no internet connection. That is what I set up in September 2026. And rather than writing everything myself, I started from an open source program that already did most of it. Here is why, how, and what it taught me.
Context
This work extends my study of a LoRa radio network for sites without coverage: a radio that only carries short text needs, somewhere, a knowledge source able to answer in three lines. The same need exists in a company: an internal knowledge server that keeps working when the connection drops, and where documents are sent to no one.
Problem
- Scattered building blocks: offline encyclopedia, course platform, book library, local language model, search engine. Each one installs, configures and updates in its own way.
- An assistant that makes things up: a language model on its own answers confidently, even when it does not know. It has to rely on the documents, and say so.
- A modest machine: no large graphics card, but an integrated graphics chip that must not be pushed beyond what it can take.
- The temptation to rewrite everything: I had started assembling my own stack, with my own search index.
Solution
1. Start from existing open source software
Project NOMAD, published by Crosstalk Solutions under the Apache 2.0 licence, is an offline knowledge server: a console that installs and updates, in containers, an offline encyclopedia (Kiwix), a course platform (Kolibri), maps, tools, and an assistant with document search. It already did most of what I was building. I dropped my own search index to use theirs.
2. Add without cutting ties with the original
I made a "plus" version of it, with one simple rule: the main branch stays an exact mirror of the original project, and my additions live on a separate branch, reapplied on top at every update from the author. My building blocks (a public portal, a cross-source search, the assistant) are separate applications, declared in the console, not code mixed into theirs. The few changes to their code are thin, one per commit, all listed in a file. The original licence is kept.
3. An assistant that cites its sources
The assistant answers from local documents: books, guides, courses. The portal, available in three languages, also offers voice, to ask a question aloud and hear the answer.
4. Measure with a fixed set of questions
I wrote 30 reference questions, asked of every version. The score rose from 18 to 28 correct answers out of 30 as fixes were made. It is this question set that measures every change, not the impression left by a demo.
Result
The server provides 68 offline collections, encyclopedias and references included, more than 25 million articles in total, a course platform, a book library and an assistant that searches about 66,000 document extracts. A daily backup is in place, and the restore procedure has been tested, not just written. The project has been paused since mid-September; my additions are not published yet.
What I take away
- Starting from open source software saves weeks. The console, installation, updates and interface were already there. My work focused on what was really missing.
- The price of freedom is discipline. Thin-layer additions, listed and replayed at every update: that is what lets you follow the author instead of drifting apart. One of my fixes is ready to be offered to them; a second is still waiting to be proven.
- Do not put everything into the AI's index. The encyclopedia made up 91% of the meaning-based search index, even though it already has its own full-text engine. I took it out of the index: the assistant searches books and guides, and the encyclopedia is consulted directly.
- A small machine has its limits. The integrated graphics chip froze several times. My first explanation, the length of the requests, was disproved by measurement, and the cause is not yet established. I separated index computation from the model that answers, so that a failure of one does not take down the other.
- A model that is too small is not enough. On a processor alone, a 3-billion-parameter model answered poorly and a 7-billion one too slowly.
- Content has a licence. I set aside one course source over a rights issue. Offline does not mean outside copyright.
Going further
Do you want an assistant that knows your documents, without sending them to a vendor, and that works even without a connection? I design AI hosted on your premises, built on open source software. See sovereign AI, my article on local LLMs, or book a 20-minute discovery call.
Sources
- Project NOMAD, official repository (Apache 2.0 licence), and project website, accessed 15 November 2026.
- Apache License 2.0, Apache Software Foundation, accessed 15 November 2026.
- Kiwix and Kolibri, official sites, accessed 15 November 2026.
- Photo: Guywelch2000 / Wikimedia Commons (commons.wikimedia.org), licence CC BY 4.0.
Tags: offline knowledge, local AI, RAG, Project NOMAD, Kiwix, Kolibri, open source software, fork