bartdegoede

It’s gets much larger (there’s the potion-multilingual-128M model, distilled from bge-m3 which has like 100 languages), so it gets to ~120m range for the lookup table (4x smaller if you’d do int8 quantization, so you could get to to 30-35mb).

Chrome ships with one (https://news.ycombinator.com/item?id=43973502) but AFAICT it’s only used for history, there’s no api for it. There’s also an included LLM? https://www.thatprivacyguy.com/blog/chrome-silent-nano-insta...

A few years ago I wrote a post about implementing tf-idf in 150 lines of python. The data source it relied on was removed (turns out Yahoo! isn't as relevant anymore).

Adventures with writing a novel with LLMs

Added an MIT license. Have fun :-)

The idea was to illustrate the concepts of an inverted index and tf-idf :-) I've built some stuff like this for a smaller project written in Python because it was Easier™ than spinning up an Elasticsearch cluster (ie it definitely didn't need to scale :-D)

Author here: you may want to check out something like Whoosh https://whoosh.readthedocs.io/en/latest/intro.html (it's like a clone of Lucene but in pure Python). I've used this to build some basic search for a small Python website and it was more than fast enough for my purposes :-)

Mailinabox (https://mailinabox.email/guide.html#admin) uses it to get SSL certificates too; you can see it in action here https://github.com/mail-in-a-box/mailinabox/blob/master/mana...

It will contain all content for pages that are live (in my case that's only 2 so far: http://bart.degoe.de/js/search/index.json), so your user will download all the content you want to have available for search on your site on page load, and then, depending on your index configuration, load more data in memory.

Lunr has an analysis pipeline that will generate a bunch of tokens for queries to match on, and you can do pretty much anything there.

It won't scale to thousands of pages, but I don't have thousands of them anyway :-)