A glossary built to be quoted

Why a glossary, of all things

Because of how the answer gets built. When an assistant answers a question about, say, chunking, it does not read a page on chunking — a retriever splits documents into passages, embeds them, and hands the model back the few that sit closest to the question. The model answers from those passages.

Which makes a definition a very particular kind of writing. It has to survive being cut out of its page and read alone. Anthropic's write-up of contextual retrieval names exactly this failure — their example chunk, the company's revenue grew by 3% over the previous quarter, is useless in isolation because it names neither the company nor the quarter. Prepending context before embedding cut their retrieval failures by 49%, and by 67% with reranking.

A glossary entry is that lesson taken to its conclusion: a passage that carries its own subject, by construction.

What is actually in it

As of 11 September 2026: 1,070 terms, and 1,590 published translations across the five languages this site publishes in. The distribution is uneven on purpose — English is complete because it is the source, and the other four are filled in term by term rather than dumped through a batch translator.

Five languages, and one that will never be here

The site publishes in English, Ukrainian, Spanish, French and Polish. Russian is not one of them and will not become one. That is a decision, not a gap in the roadmap.

It is also enforced mechanically rather than hoped for, because good intentions did not catch it: a hand-authored Ukrainian entry once shipped with a Russian word inside it, typed by somebody writing five languages in one sitting. A script now scans anything hand-written in Cyrillic for the four letters that exist in Russian and not in Ukrainian — ы, ъ, э, ё — which is a high-precision signal, since none of them can appear in correct Ukrainian. A reading agent catches the rest: the endings and the false friends that use only shared letters.

Every generation prompt names its target language and forbids Russian explicitly, because a model asked for Ukrainian produces Russian words inside it often enough that saying so is worth the line.

How an entry is made, and what stops a machine touching it

Most entries are drafted by a model, then reviewed, and published only when both approval and publication are set — nothing a model produced is public on the strength of having been produced. Each translation stores which model drafted it, and that field is not decoration.

The guard is per term, not per language. A term with hand-written work in any language is hand-maintained, because a queue that quietly filled in the missing languages would leave one page in considered prose and the next in draft. Handing an entry back to the queues means clearing that field — a decision somebody makes and records, not a checkbox on a batch screen.

The marker is the record of how the entry was made. The thing that records how it was made is the thing that decides whether a machine may touch it.

Why the definitions are short

A long definition is a worse definition for this purpose, and the reason is mechanical rather than stylistic. If the explanation runs across four paragraphs, a retriever takes one of them. The one it takes may be the caveat, or the history, or the example — and the answer built from it will be confidently incomplete.

So each entry leads with a definition that stands alone: what the term is, in a form that survives being the only thing anybody reads. The depth comes after, where a reader who wants it will find it and a retriever that grabs it will still be holding something true.

Terms that depend on each other are linked, so a reader who lands on one can reach the one it assumes. That linking is also what makes the glossary a structure rather than a list.

What it is for, in one sentence

If an assistant is going to define your industry's vocabulary to your buyers, the definitions should be written by somebody who had to be right. That is the whole ambition: a reference short enough to be quoted correctly, complete enough to be quoted often, and honest about which sentences a machine wrote.

If you want to see what a retriever would take from a page of your own, Vectorscope shows you the passages. If you want the instruments that produced this vocabulary in the first place, they are all on the free tools shelf.