How the search works

Updated 2026-10-17 (UTC): when your words reach the address bar, and what an older browser shows. This note is about where the words you type into the search box go, and what your browser keeps of them. From 2026-10-15, when the search went live, until 2026-10-17 (UTC), the page wrote the box's words into the address after #q= each time you paused while typing, and Chrome and other browsers built on Chromium kept each version as its own history entry, words you deleted included. From 2026-10-17 (UTC) the words enter the address only when you press Enter, open a result or press the page's "link to this search" button, and they leave it again when the words in the box change, so a browser built on Chromium keeps one history entry for each of those acts instead of one for every pause. The page also checks, before it shows the box, that your browser has what the search needs: a browser older than the script's own patterns (before Chrome 64, Safari 11.1 or Firefox 78), or one that fails that check, shows no box at all, only the message for readers without JavaScript and the full list of articles, which find-in-page searches. Set against this page as published on 2026-10-15: the two things it said a later update would fix, the address rewritten at every pause and a box that showed but did nothing in older browsers, are fixed, and the sentences below that described them are corrected in place.

1 · What it is, and where to find it

The search lives at /search/. It lists every article on this hub, newest first: 82 when it went live on 2026-10-15, counting two from the diffusion wing, our image-generation section: its front page and one bench. Type in the box and the list narrows and reorders; from a result, press / or Esc to come back to the box. Under each result, a why line says which of your words matched and where: the title, the summary, a section heading, the short version, and how many times in the text. Without JavaScript, or in a browser too old for the search, the page is still the whole catalogue, and your browser's find-in-page searches it. A mistyped address on this site lands on a not-found page with a search box of its own; Enter there opens this search with the words you typed into it.

2 · How to use it

These were typed on the live page on 2026-10-15, before this page joined the index; as articles arrive, the counts and the order move.

• A graphics card: 3090. 32 articles match. First is "Two 3080s against one 3090, and the cap decides", whose why line reads: 3090 in the title, the summary, a section heading and the short version, and 65 times in the text. A word with a digit is never folded and never guessed at: 3090 never meets 3080. A card name typed with a hyphen where an article prints a space (RTX, a hyphen, 3090) still finds the articles that print RTX 3090, and the count line says so.
• Two engines (programs that run AI models): ollama vs vllm. 15 articles carry every word and 31 carry some; "Ollama or vLLM: which to run, and how" is first. Small words like vs, the and how are skipped.
• A model tag (a label for how a model file is compressed): Q4_K_M. "The compressed photograph — what those Q4_K_M tags actually mean" is first. A tag is searched whole and by its parts, and llama.cpp also meets llamacpp.
• Another spelling: quantisation. The count line adds "also searched: quantization", and the compressed photograph is first again, on 12 times in its text.
• A plain question: how much electricity does a gpu use. 2 articles carry every word and 72 carry some; "What it costs to run a high-end AI rig at home" is first, and the articles holding only some of your words follow below a divider.
• A slip: 5900, the 5090 with two digits swapped. The page says nothing in the index matches it and shows its standing examples, the same for every miss; it does not guess what you meant.

3 · How it works

One file. When you first use the box, or open a link to a search, the page fetches one index file from this site, /data/search.json: about 0.3 MB compressed, the same bytes for every reader. It holds each article's title, summary, section headings and short version, and every word of its text but the small ones the search skips, with how often it appears, but not the text's sentences. From then on, your browser does the work.

How a page is scored. A short table folds spellings into one (licence and license); on longer words a plural, -ing or -ed folds away, so cards meets card; a word with a digit is never folded. Each article is scored by BM25, a standard ranking formula: a word counts for more the more often an article says it and the rarer it is across the site, and for more in a title, summary or heading than in the text. Articles with every one of your words come first, then those with some; a tie goes to the newer article. The last word you type also finds longer words that begin with it, at less weight, and only in an article that does not hold the word itself.

Where your words go. The script's only request is the index file, and nothing about your words is in it. While you type, your words stay in the box. When you press Enter, open a result or press "link to this search", they go into the address after #q=, and browsers leave out everything after a # when they ask a server for a page, so this site sees only that the search page was opened and the index fetched. Our visit logs keep request paths for thirty days, as the privacy page says; a search leaves no word of it there. Opening a result asks this site for that page, as any link does, with no referrer (the note of which page you came from), because the site's header tells browsers to send none. The search itself stores nothing: no cookie, nothing in your browser's storage, and the box is marked so the browser saves no form entries from it; spell-check is off in the box, so a browser whose spell-check runs online is not handed the words. No other site: the page fetches nothing from any other host.

A test holds that path: a browser run by a script (headless Chromium) opens a built copy of the site, searches two made-up words and opens a result; the requests must be exactly the search page, the index and the result, with neither made-up word in any of them. It checks the requests for those queries on that path, not your browser's history or every query a reader could type.

What your browser keeps. Once your words are in the address, your browser's history keeps that address, and a browser that syncs its history or open tabs to an account can send it there. The page puts them there only on Enter, on opening a result or on "link to this search", and takes them out again when the words in the box change, so Chrome and other browsers built on Chromium keep one history entry for each of those acts, not one for every pause (we have not measured Firefox or Safari). If your browser translates this page, its built-in translation can send the page's text to its translation service: the lines that repeat your words are marked not to be translated, but the results, with the words that matched highlighted, are translated like the rest of the page. Anyone you share a search link with can read the words, and so can the app you share it through.

Two pages now run a script. Until 2026-10-15 no page this hub publishes itself ran a script (the diffusion wing, built from another of our repositories, has its own theme switch). Now two run one hand-written file of ours each, inline, with no library: the search page and the not-found page. At every deploy, the check that publishes the site refuses a script on any other page that opens without a password, and on these two any script that is not the reviewed file, byte for byte. Four published sentences that said otherwise, two in llms.txt (the site's machine-readable map) and two on the privacy page, are corrected in place, dated 2026-10-15.

4 · What it cannot do

• Meaning. It matches words: a question that uses none of an article's words does not find it.
• Guessing. Beyond the spelling table and the endings it folds, it guesses nothing: a misspelt card name finds nothing, and the page says so rather than suggest one (the examples it shows are the same for every miss).
• Popularity. No ranking by popularity, no record of what anyone searched, no suggestions.
• Two pages' text. The two diffusion-wing pages are in the index by title and summary only; the wing's other pages are not in it.
• Older browsers. A browser older than the script's own patterns (before Chrome 64, Safari 11.1 or Firefox 78) runs none of it, and one that lacks what the page checks for before it shows the box never shows it: either way there is no box, only the message for readers without JavaScript and the full list, which find-in-page searches.
• A few held-out words. Words the check that publishes this site would refuse standing alone, out of the sentence that makes them safe, are left out of the index's text (two at the live build); a title, summary, heading or short version still finds them.

5 · How we checked it

A ranking is a set of choices, and a choice can be wrong, so the search carries a gate: 24 queries a reader might type, each with what must be true of its results (say, which article comes first or sits in the top three, or what the why line names), run on a frozen copy of the index, all or nothing, on every change to the ranking. The first 22 were frozen on 2026-10-13 at 14:39Z, before the ranking code was written; two were added at 18:24Z the same day, after two faults we planted in the ranker were caught by no row. A row changes only by a dated amendment on our decision; amended so far: 0. The limit: we chose those 24 queries with the index in front of us. A sealed set, written by someone who never saw the ranker, waits for a later version.

How to check our work

Open /search/ with your browser's network panel showing and type anything: you will see the page and the index, then nothing until you open a result; your browser's history page shows the addresses it kept. The page's source holds one bare <script>, beside one type="application/ld+json" data block for search engines, which a browser does not run. /data/search.json is the whole index, open to read.

The rest of the seminar

• Where your question goes: the privacy page, in plain English, with the dated note this search required.
• llms.txt: the site's machine-readable map, corrected the same day.

Who ran this, and thanks

The ranking code, the tokenizer and the index builder are the workshop's own, built on no library. The checks behind them stood on Chromium (run headless), Node.js and Python's standard library, none of which owed us anything. A small (human) team decided what the search would and would not do; a fleet of AI agents wrote the scripts, ran the tests, audited them, and read these pages as a stranger would.