Document pipeline

Glossa is built around a four-phase document workflow:

1. Configure the translation pipeline, glossary, and language pair.
2. Test one chunk first so you can inspect output without locking the whole document.
3. Translate the full document once the setup is stable.
4. Review the audit results and iterate if the quality needs work.

The underlying reasoning is explained in LLMs and pipelines: a language model is powerful but probabilistic, so Glossa splits work into chunks, stages, and audit passes to keep each step controllable.

Document workflow

Glossa uses the document workflow even for short trials. The path stays the same: create or open a project, import or paste the source text, review chunking, and use Test on one representative chunk before starting a batch.

Area	What you control	When to use it
Import and preview	Source text, segmentation, initial chunks	Before creating the active chunk list
Pipeline config	Languages, providers, models, prompts, glossary	Before Test and before long batches
Document view	Current chunk, output, states, runs	During translation and review
Insights panel	Audit, notes, statistics, coherence	During quality control and chunk closure

DeepL Hybrid Mode

The DeepL Hybrid mode combines the speed and precision of the DeepL API with the contextual refinement of an LLM:

Stage	Provider	Role
Stage 1	DeepL API	Main translation
Stage 2	Optional LLM	Style and register refinement
Judge	LLM	Quality audit (unchanged)

Requirements: DeepL API key configured in Settings → provider section.

When to use it: Texts that require high terminological fidelity and speed, where an LLM alone would need too much context or elaborate prompts.

Formality: For languages that support it (German, Italian, etc.), you can set the formal/informal register directly in the DeepL stage.

DeepL Glossaries: You can build a DeepL glossary from the terms in your Glossa glossary assigned to the pipeline, so DeepL automatically respects your terminology.

Standard document flow

1. Import a document or prepare a short sample.
2. Choose chunking and confirm the import preview.
3. Set the source and target languages.
4. Choose the provider and model for each active stage.
5. Add glossary entries or phrase-memory retrieval if the project needs terminology control.
6. Run a test chunk.
7. Review the candidate translation, audit findings, and chunk metadata.
8. Switch to production mode and process the remaining chunks.
9. Lock or revise chunks as you complete editorial review.

Run state	Purpose
Test	Preview one chunk and keep the config editable
Production	Process all remaining chunks

Stage behavior

Stage	Purpose
DeepL Translation	Produces the first translation through the DeepL API when DeepL Hybrid is active
Translation	Produce the first candidate from the source chunk
Refine	Rewrite the candidate with better style, accuracy, or terminology
Format	Clean the output format without re-translating the source
Judge	Evaluate the result and report structured issues
Coherence	Check consistency across translated chunks when enabled

Editorial mode exposes more of these stages explicitly. Standard mode keeps the workflow lighter.

What you review at each phase

Phase	Main question
Configure	Are languages, stages, prompts, and glossary correct?
Test	Is one representative chunk good enough to scale out?
Translate	Is the batch progressing cleanly and producing stable chunks?
Review	Which chunks still need editorial intervention?

What survives between runs

• Completed chunks are not recomputed unless you explicitly re-run them
• Cancelled batches resume from already completed work when possible
• Test runs do not lock the whole configuration
• Review data and annotations stay attached to the chunk they describe

Common mistakes

• Switching provider and prompt at the same time, then not knowing what changed
• Moving to Production before one difficult chunk has passed Test cleanly
• Using the format stage to repair translation errors instead of formatting only
• Treating chunk completion as final acceptance without reading the audit

Practical rules

• Stay in Test until the prompt, glossary, and model choice stop moving.
• Use Production only when you want the remaining document to follow the same setup.
• If the formatting stage starts changing meaning, remove or simplify it.
• If a chunk is difficult, annotate it instead of relying on memory alone.

See also

• Annotations — for tracking editorial findings per chunk
• LLMs and pipelines — why Glossa separates chunks, stages, and audit
• Audit and review — detailed review loop with the judge
• Context and caching — how Glossa uses context across adjacent chunks
• Pipeline config — full reference for all controls