Glossary, translation memory and brand voice

These three belong to the organization and apply to all its sites and adapters.

Glossary

Property	Behaviour
Scope	Per organization and source language.
Term	Up to 200 characters, unique per source language.
Targets	One translation per target language, up to 200 characters. Languages without a target have no rule.
Do not translate	The term stays unchanged in every language.
Case-sensitive	Only the exact spelling matches. Otherwise case is ignored.
Notes	For editors; not sent to engines.
DeepL	One DeepL glossary for your organization with an entry list per language pair, updated when terms change. Pairs without DeepL glossary support get none. If the glossary can’t be prepared, the language’s segments fail instead of being translated without it. Exception: if DeepL’s glossary limit is reached, the language is translated without the DeepL glossary and the job says so; the QA check still flags terms that weren’t applied.
LLM	Only terms that occur in the text (whole words) are added to the instructions.
QA	Do not translate violated: error. Target term missing: warning.
Count	glossary_terms in GET /status.

Translation memory

Property	Behaviour
Scope	Per organization and language pair, across all sites and adapters.
Entry	One per source text (by fingerprint) and language pair. The latest approval wins.
Learns from	Approvals reported through POST /reviews (with the editor’s final text), imported translations, TMX imports, and edits in the dashboard. Approvals of translations whose source has changed since are skipped.
Exact match	Same source text after normalization, same format and language pair. Applied before an engine is chosen; delivered as a draft with origin tm_exact, engine tm. Not billed.
Fuzzy match	Similar, but not identical, source texts are given to the LLM as reference translations. Never applied automatically, and not used with DeepL.
Repeats	The same source text several times in one job is sent to the engine once. The copies are not billed (reused_segments, reused_characters).
Review	TM hits are still delivered as drafts and reviewed by editors.
TMX	Import and export in TMX 1.4. Import flattens inline markup; files up to 10 MB.
Deleting a site	TM entries stay with the organization.

Brand voice

Property	Behaviour
Levels	Organization → site → site + target language.
Formality	formal, informal or not set. The most specific level that sets one wins. The site’s language setting (Formal/Informal) takes precedence; Default there means “use the brand voice”.
Instructions	Free text. All levels apply, general first.
Forbidden words	All levels apply together.
DeepL	Receives the formality only, where the language supports it.
LLM	Receives the formality, the instructions and the forbidden words.
QA	A forbidden word in the translation is an error.