Translating Three Hundred Pages Without Renaming the Equipment

Technical documents are not written like prose. A safety case, a tender, an equipment manual — each one leans on a fixed vocabulary: the same terms, the same abbreviations, the same equipment and process designations, used identically from the first page to the last. That repetition is the point. A term that means one thing on page 4 has to mean exactly that on page 380, or the document stops being a technical document.

They are also long. Three hundred pages is ordinary; a thousand is not rare. No model translates that in one pass. The text has to be cut into pieces — often hundreds of them — and the pieces get translated independently, which is where the trouble starts.

An awkward sentence is survivable. A valve that appears under three different names is not.

What Goes Wrong

Set an AI to translate a specialised term it does not know and it will do something reasonable: invent a rendering, expand the abbreviation, or translate it letter by letter. It is not being careless. It has been handed one chunk and asked to translate it well, and inside that chunk its choice is defensible. It simply has no way of knowing that the same word was settled elsewhere in the same document.

Multiply that by the number of decisions a technical document contains and the result is not a translation. It is a set of translations that happen to share a file name.

The term defined in the specification no longer matches the term used in the body. An abbreviation that appears fifty times resolves into a dozen different phrases — sometimes expanded, sometimes not, occasionally translated as though it were an ordinary word. Equipment designations that should have been left exactly as they were come back translated. A cross-reference points at a heading whose wording has changed underneath it.

A reviewer can no longer tell whether two names mean the same part or two different parts, which in a technical document is the only question that matters. An engineer searching the translated manual for the term used in the specification finds a third of it. If the document is going to a regulator or a certification body, inconsistent terminology is not a blemish on the style; it is a finding.

None of this looks like a translation error. Every sentence reads fine. That is what makes it expensive: the document passes a spot check and fails a review.

Two Ways to Run It

This is the problem the Document Translator in the Corporate AI Portal was built around, and it is why the module has two modes rather than one. Holding terminology steady costs tokens, and not every document needs it.

The first mode translates straight through: the text goes out in pieces and comes back in pieces. There is no glossary in this mode at all — those agents are never given the terminology tools, which is a large part of why it is cheap. For an email, a memo, a slide deck or a spreadsheet of part numbers that is exactly the right tool: quick, cheap, predictable. It has one limitation on a technical document: no memory of itself.

The second mode adds a terminology layer. Agents can look a term up, record a term they have settled on, and check a draft against what has already been decided. It costs more tokens per page, because those lookups and checks are extra work the model has to do. On a document where the vocabulary carries meaning, that is the difference between a translation you can send to a reviewer and one you have to retype.

The Glossary Belongs to the Document

The mechanism is the oldest one in translation: a glossary. What is new is that nobody has to prepare it.

Normally a term base is agreed before the job. Someone reads the document, pulls out the terminology, decides the renderings, and hands the list to the translators. On three hundred pages of a narrow specialism, that someone does not exist — and even when they do, the document is usually already late.

So the glossary assembles itself while the translation runs. Agents are given three tools, the three any translator would ask for:

• Look up. Has this term been handled already, and how? The answer comes back from the table, or comes back empty.
• Record. I have settled this term. Store the pair, so that nobody has to decide it again.
• Verify. Check my draft against what the document has already agreed, and show me where I have drifted.

Agents call them as they go and the table fills up behind them. Nobody has to decide in advance which terms will matter: the ones that matter are the ones that recur, and those are exactly the ones an agent will look up.

Deciding what an abbreviation stands for is the model’s job. It is a judgement about meaning — is this “RPV” here the same thing it was on page 12? — and judgement about meaning is what these models are good at. Making the answer stick across three hundred pages is not a judgement call at all; it is bookkeeping, and bookkeeping is better done by a table. Each side does what it is suited to.

This split is where the consistency actually comes from, and it is worth being precise about why. The model never sees the document. It sees one chunk at a time and forgets the rest, so asking it to keep a vocabulary consistent is asking it to remember something it was never shown. The table sees every decision that has been recorded, in order, and forgetting is not something it does.

The glossary belongs to the document, not to the agent. Every agent reads and writes the same one while up to four of them run at once. When the agent on page 57 meets the containment vessel, it looks the term up and gets what page 3 decided. The term is no longer a series of local guesses; it is a decision the document holds.

Definitions and abbreviations gain the most from this. The paragraph where a document defines its terms is the anchor for everything that follows, and previously that anchor reached no further than its own chunk. Now the first agent to meet an unfamiliar abbreviation works out what it means from the sentence around it and records the result, and every later occurrence, in any chunk, resolves the same way. The reading experience this produces is unremarkable, which is the aim: a translated manual should no more surprise you with a renamed part than the original would.

There is a practical reason the glossary is reached through tools rather than pasted into the prompt. A term base for a thousand-page document does not fit in a context window, and it is still growing as the translation proceeds. Asking for one entry at a time costs a few tokens; carrying the whole table in every request does not scale, and a model that cannot see the whole table will start improvising exactly as it did before.

Your File Comes Back as Your File

Terminology is half the problem. The other half is the file. A translated manual is not much use if the tables have collapsed and the diagrams have moved.

So the translator does not rebuild your document. It opens the original and replaces only the text. Fonts, styles, images, table layout, slide geometry and cell formatting are carried through untouched, because they are never copied out or reconstructed in the first place — and what was never touched cannot be got wrong.

Some content is deliberately left alone. Spreadsheet numbers stay numbers: turn a number into text and every sum below it quietly skips that cell. Formula results stay where they are, because the value stored beside a formula is the machine’s own answer, not something the author wrote. Mathematical notation is lifted out before the model ever sees it and put back afterwards, so an equation cannot be “improved” in translation.

Markdown and CAD drawings are written out from what was captured rather than patched in place. A drawing declares its own text encoding, and that declaration is what gets used, so a Western European or Vietnamese drawing reads the way its author wrote it rather than as mojibake.

One honest exception: PDF. A PDF has no paragraphs to preserve, only glyphs at coordinates. The PDF path is a conversion into a Word document, and tables, images and layout do not survive it. If you have the source document, send the source document.

What It Does Not Do

The glossary lives with the document. It does not travel between documents yet: translate the same equipment manual again next month and the terminology work starts over. A term base that accumulates across a whole programme is a different piece of work, and a worthwhile one.

Andrey Obushev