Pipeline configuration
Configuration belongs to a project’s pipeline. Service credentials and connections belong to application settings.
Sections
| Section | Parameters |
|---|---|
| Settings | Mode, languages, persona, examples and general options |
| Translation | Providers, models, prompts and generation-stage options |
| Quality Control | Evaluator and consistency check |
| Term registry | Assigned glossary |
| Prompt Preview | Request structure for active stages |
Standard, Editorial and DeepL Hybrid modes are described in the translation workflow.
Languages and persona
Set source and target languages. A persona is free text that replaces the default opening of the system message. It can specify role, subject area, languages and register. When enabled, it should state the intended language pair and instructions accurately.
Prompts can be saved as reusable templates, organised by context. Prompt refinement sends the current text to a configured model and places a revised version in the field. It requires a connection and any credentials needed by the selected provider.
Model parameters
Each LLM stage selects its provider and model independently. Available controls depend on capabilities declared in the application.
- Temperature: controls sampling variability without guaranteeing accuracy or repeatability. Supported ranges are 0–1 for Anthropic and 0–2 for Gemini, OpenAI and DeepSeek.
- Reasoning: when requested for OpenAI or DeepSeek, the adapter omits the temperature parameter.
- Ollama: provides context, generation and reasoning options according to the model. Advanced options must be a valid JSON object, such as
{ "num_ctx": 8192 }. - Ollama assessment: schema-constrained requests set temperature to zero, even when another value is configured.
Invalid JSON does not replace the previous configuration. Check the meaning of service-specific options for the model you use.
DeepL Hybrid
The first stage uses DeepL settings: language, formality where supported, translation mode and remote glossary. Its API key is separate from the keys used by revision and assessment LLMs. DeepL quota or glossary errors must be resolved with that service before the sequence can complete.
Examples and context
A pipeline can hold up to five translation examples, added from a locked segment’s Audit tab and edited in settings. Phrase memory instead supplies references selected for an individual segment. Prompt caching has provider-specific rules.
Cost estimates
The configuration estimate covers the entire document, including consistency review when configured. In the document sidebar, the estimate follows the selected action. Details distinguish stages and models.
Estimates use an approximate word-to-token conversion and the model prices recorded in Glossa. Usage shown after execution uses token counts returned by the service; the associated cost is still Glossa’s calculation, not an invoice. DeepL reports billed characters, which are not LLM tokens and are not included in the same dollar estimate.