Skip to content

Pipeline config

Glossa separates pipeline configuration from document content so you can tune the run before starting a batch.

If you want to understand why controls are split by stage, also read LLMs and pipelines: it explains why translation, refinement, formatting, and judging have different responsibilities.

Main controls

  • Source language
  • Target language
  • Pipeline mode
  • Provider and model per stage
  • Translation instructions
  • Persona
  • Glossary / term registry
  • Phrase-memory settings
  • Audit / judge settings

Typical configuration surfaces

SurfaceWhat you usually set there
Settings tabLanguages, run mode, general defaults, persona
Translation tabStage prompts, models, provider-specific options
Audit tabJudge model, judge prompt, coherence prompt
Glossary areaAssigned glossary and term entries

What usually changes first

If a run is not good enough, change these in order:

  1. Translation prompt
  2. Provider or model
  3. Glossary entries
  4. Phrase-memory retrieval
  5. Judge prompt

Leave everything else alone until you know which part is causing the failure.

Pipeline modes

ModeDescription
StandardSingle translation pass plus audit
EditorialTranslation, refine, and formatting stages before audit
DeepL HybridDeepL first pass, optional LLM refinement, and LLM audit

Stage-level advice

  • Keep the translation stage focused on accuracy and basic style.
  • In DeepL Hybrid, use the DeepL stage for the first draft and keep LLM prompts/models separate for refinement and judging.
  • Use refine for rewriting, not for first-pass translation.
  • Keep format narrow so it does not silently alter meaning.
  • Use the judge to report issues, not to replace human review.

Execution rules

  • Test mode processes one chunk and leaves the configuration editable.
  • Production mode processes the full document.
  • A cancelled run resumes from the already completed chunks when possible.

Stability advice

  • Change one major variable at a time.
  • Save the pipeline before large batch runs.
  • If a project is stable, clone or rename a pipeline before experimenting.

When to change config

Change the configuration before a full run if you need a different provider, prompt, or glossary behavior. If you only need to inspect a result, prefer Test mode over changing the whole pipeline.

Cost estimate

Hovering over the info icon next to the estimated cost (in the pipeline settings panel, and next to the translate/run button in the document view) shows a per-stage breakdown with the approximate cost in dollars.

  • In the pipeline settings panel the estimate always covers the whole document, including the coherence check if configured.
  • Next to the run button in the document view, the estimate follows whatever is about to happen: in "translate chunk" mode it covers only the selected chunk, in "run all" mode it covers the whole document.
  • This is an approximation based on word count and the chosen model's per-token price: the real cost may vary slightly.
  • DeepL stages are measured in billed characters by DeepL: Glossa can show those after the run, but the dollar estimate remains based on token-priced LLM providers.

See also

Public documentation for the Glossa desktop app