Knowledge Base Translation at Scale

Quick answer — At knowledge base volume, translate on the change rather than in batches and let segment memory absorb the repetition. The cost driver is churn, not article count.

Vitra.ai Universe ships product and customer content in every language.

Volume is not the problem, churn is

Two thousand articles sounds like the challenge. It is not — a one-off batch of two thousand is a project with an end. The problem is that several hundred of them change every quarter. Batch translation means either paying to retranslate whole articles for a one-paragraph edit, or letting the translated version drift out of date, and most teams end up doing both.

Translate the change

Segment-level memory is what makes this work. An edited paragraph is retranslated; the rest of the article is an exact reuse and costs nothing.

That changes the economics enough that keeping forty languages current becomes ordinary rather than a budget decision — which matters because a knowledge base that is only current in English quietly trains everyone to read English.

Retire properly

The last row is worth dwelling on. Knowledge bases accumulate articles about features that no longer exist, and every language multiplies that.

Unpublish rather than translate. An article nobody should read is not improved by being readable in nine languages.

Structure beats prose

Articles built from consistent patterns — a stated problem, numbered steps, a result — translate more cheaply and reuse more, because the scaffolding repeats across hundreds of articles and memory recognises it.

Free-form prose does neither. That is a writing standard rather than a localization one, and it pays back most at this volume.

Where a person still belongs

Billing, security, data handling and anything describing a contractual commitment.

Everything else publishes once quality control clears it, which is the only shape in which a small team can keep pace with the release cadence.

Watch lag rather than volume: how many hours behind English each language runs. It is the number that says whether this is working, and almost nobody tracks it — the argument in continuous localization.

FAQ

What makes knowledge base translation expensive? Churn rather than article count. A one-off batch has an end, but several hundred articles change every quarter, and retranslating whole articles for one-paragraph edits is where the budget goes.

How do you avoid paying to retranslate unchanged text? Segment-level translation memory. Only the edited paragraph is retranslated while the rest of the article is an exact reuse, which makes keeping many languages current an ordinary operation.

What should happen to outdated help articles? Unpublish them rather than translate them. Knowledge bases accumulate content about features that no longer exist, and every additional language multiplies that maintenance problem.

What should be measured in a large knowledge base? Lag — how many hours behind the source each language runs. Volume translated says nothing about whether readers in that language are seeing current information.