This could be helpful even for people who already have internet access. The internet is full of distractions, and we often end up seeing ads instead of actually learning something. It would be great to create personal collections of useful websites and have a powerful search tool to explore them—without relying on Google and getting sidetracked.
Right now, Kiwix’s search function is quite basic and doesn’t work well with large amounts of content. It might be worth exploring the use of a generative language model or embedding model to improve its accuracy and usefulness.
These devices are insanely underpowered for running LMs. We've had pretty decent search long before the current hype wave. Postgres has built-in support for FTS in multiple languages[1]; if you want to get fancy you can maybe throw in a table of common synonyms.
This is a solution that is supposed to be affordable in third-world countries, where 99% of all personal computing happens on low-to-mid-end smartphones.
Unless you specifically mean first-world countries, in which case you can just run a hotspot from your PC or laptop.
> it is how easy or hands off the appliance can be.
I can already barely justify having a single RasPi plugged into my router, and I'm a devops nerd. That box does one single thing that is too annoying to set up on my Mac mini. 99% of people do not want yet another blackbox appliance, this is why every single SOHO router has a built-in switch and AP.
> Every postgres search implementation has atleast one internet connected postgres nerd on standby.
It doesn't have to be postgres specifically, just pointing out that decent FTS has been commodity software for decades. Throwing LMs at every already-solved problem is the problem with LMs.
I agree with you on every step and every point except the LM. Your entire argument hinges on FTS requiring a fat RDBMS. This is not the case.
Postgres is what I have experience with, but I just looked up SQLite and indeed it has FTS5[1]. This is not an LM-grade problem, this is a solved problem.
Right now, Kiwix’s search function is quite basic and doesn’t work well with large amounts of content. It might be worth exploring the use of a generative language model or embedding model to improve its accuracy and usefulness.