Glossary

The Glossary is a list of source-language terms paired with their preferred target-language renderings. Sublarr injects matching entries into the LLM prompt for every translation job, so a name that gets translated one way in episode 1 doesn’t drift in episode 14. There are two scopes — global (applies to every job) and per-series (applies only to that series).

When to use it

Scenario	Add to glossary?
Recurring character name that the LLM pronounces inconsistently	Yes — per-series
Place name with multiple valid translations	Yes — global
Idiomatic phrase that needs a fixed rendering	Yes — global
One-off line that came out wrong	No — fix in the editor instead
Whole sentence transformations	No — glossary is per-term, not per-line

Where it lives

Surface	Scope	When applied
Settings → Translation → Glossary	Global	Every translation job that has glossary injection enabled.
Series detail → Translation → Glossary	Per-series	Only jobs for that series. Layered on top of the global list.
Subtitle Editor → Translation tab → Glossary toggle	Per-job	Disable glossary for a single run without removing entries.

Entry shape

Each entry has three fields:

Field	Effect
Source term	The string the LLM will look for in the source-language cue. Case-insensitive match by default.
Target term	What it must be translated to.
Notes	Free-text comment shown to admins (not sent to the LLM). Useful for “Variant of X” reminders.

Tip

Keep entries narrow. A glossary that includes “the” → “der” will drown signal in noise. Focus on names, places, recurring jargon, and titles.

How injection works

When a translation job runs:

1. Sublarr collects matching entries from the global + per-series lists.
2. The matched entries are appended to the system prompt as a “Required terminology” section.
3. The LLM is instructed to use the listed renderings verbatim, and to ignore entries that don’t appear in the cue.

The entry budget is capped (default 40) to keep prompts within context limits. If a series has more matching entries than the cap, Sublarr picks the most-recently-edited subset.

Per-series cascade

The series-level list augments the global list — it does not replace it. If a term appears in both, the series rendering wins. To explicitly override a global entry for one series, add the same source term with the new target term to the series list.

Auto-build from existing translations

The glossary supports an “auto-extract” pass that reads through your existing translated files and proposes glossary entries based on repeated patterns:

1. Open Settings → Translation → Glossary → Auto-extract.
2. Pick a series and the language pair.
3. Sublarr scans existing translated .srt files, finds proper-noun candidates, and presents them as suggestions.
4. Review, edit, and accept entries one at a time.

Caution

Auto-extract is a starting point, not a verified list. Skim every suggestion before accepting — proper-noun detection has false positives, especially for stylised anime names.

Backup and export

Both the global and per-series glossary are stored in the database and included in Settings → System → Backup. The export format is YAML for human-readable diffs:

global:
  - source: Kyoto
    target: Kyōto
  - source: Senpai
    target: Senpai
series:
  "Frieren: Beyond Journey's End":
    - source: Himmel
      target: Himmel  # Keep German as-is

Common pitfalls

Symptom	Likely cause	Fix
Term still translates inconsistently	Source-term casing differs in the cue (e.g. Senpai vs senpai)	Add both casings or enable the case-insensitive flag.
LLM ignores the glossary	Glossary toggle disabled for that job	Re-run with the toggle on.
Translation degraded after adding many entries	Prompt budget exceeded	Lower the per-job cap or split global entries into per-series lists.