Knowledge

Knowledge is a way to own a small piece of the internet. Most of what I read these days sits inside a handful of platforms whose ranking isn’t tuned for me. I wanted a site that’s openly biased toward the stuff I find interesting, with my own ranking deciding what comes first.

I started Knowledge right after my PhD. After months of reading papers and writing chapters, I badly needed to type something that wasn’t prose (touch some grass, but with a keyboard). The primary inspiration was Semanlink, a personal knowledge base my PhD supervisor at Renault quietly tends to, and has for close to twenty years.

The search stack borrows technologies we design at LightOn. The matching model keeps a vector for every token in your query instead of squashing the whole thing into one, which is why a paraphrase or a typo still finds the right page. The server runs a quantised version of the model for speed, and the browser quietly loads the full-precision copy to re-rank the top hits before they show up on screen. If you want to dig into how any of this works, the LightOn website, LightOn’s GitHub, and my own GitHub are the right places to start.

Knowledge plays well with Zotero. Drop in an API key and your reference library shows up next to everything else, fully searchable. Zotero’s mobile app makes a great phone bookmark button too: save something on the bus and it lands in Knowledge on the next sync.

Knowledge will lean further into hierarchical search and tighter LLM integration over time. Pairing that with multi-vector retrieval feels like the most promising combination for the kind of grounded, library-first search I want to build here.

Thinking about opening GitHub Sponsors so the parsing and indexing pay for themselves. Sponsors keep their sources flowing, and they can pull the cleaned, tagged data out whenever they want. Not sure yet.

A small note for anyone listed here. I’m not trying to bother anyone, just keeping a few smart people’s reading in one place. If you’d rather not have a library on Knowledge, drop me a line at raphael.sourty@gmail.com and I’ll take it down the same day.

And if there’s someone whose reading you’d love to see in here, open an issue on the Knowledge repo with their handles and I’ll add them.

I keep an anonymous count of page views, search queries, clicks, filter use and device type. That’s all allowed under the CNIL audience-measurement exemption. No IP addresses, no cookies, no fingerprints, nothing shared with anyone else.

— Raphaël Sourty · @raphaelsrty