Context retrieval
Understand how an engine selects guidance and investigate an unexpected translation against its source and configuration.
If a translation uses an unexpected term, two different things may have happened: the request did not receive the glossary entry, or the model received it and used it incorrectly. The repair depends on which happened. Context retrieval explains how saved guidance becomes part of a request.
The localization engine loads applicable instructions, selects a brand-voice variation and model chain, and prepares glossary retrieval. The selected material joins the source and any request-specific context in the translation prompt.
What each resource contributes #
| Material | How it is selected |
|---|---|
| Rules | All matching rules accumulate from attached rulesets and predefined selections, including applicable wildcard and regional guidance. |
| Brand voice | One target-locale variation wins through locale ranking and tie-breakers. |
| Glossary | Locale-safe terms are ranked, deduplicated, and selected for the source content. |
| Request context | The caller supplies content and context through its own supported contract. |
Predefined rules precede custom rules. Glossary instructions include the preferred translations and hints. This arrangement gives the model relevant guidance and expresses priorities; it does not enforce compliance independently of generation.
Why a glossary entry may not appear #
A small applicable term pool can be preloaded. A larger pool uses semantic retrieval to find relevant terms rather than sending every record in the organization. An attachment alone does not mean every term reached a particular request.
Locale ranking and source-text deduplication happen before term instructions are assembled. Script safety excludes incompatible custom translations. An entry can therefore be saved and attached correctly without appearing in every request.
A hint explains the intended sense once a term is included. It is not a hard filter that prevents all unrelated senses from matching. Use Glossaries and Locale resolution to distinguish term meaning from eligibility and ranking.
Trace an unexpected result #
- Open the run and confirm its engine, source locale, document, and target locale. A bare locale can resolve differently from the region you intended.
- Inspect the source batch and output in Requests. Compare them with the submitted source revision and document context.
- Read the engine's attachments and the relevant terms, rules, and voice variation. Creating a resource does not attach it to an engine. Check locale scope, term type, and the instruction text itself.
- Compare the available configuration history with the time the job ran. The request view shows the source batch and raw output, but does not expose a complete historical prompt. Current attachments alone cannot establish exactly which guidance an earlier attempt used.
- Correct the suspected selection or guidance problem, then translate a bounded sample and a case the change should leave intact. Save the engine settings and submitted context with the new result so the comparison can be explained.
Keep an uncertain cause separate from the observed wording defect. A better result after a guidance change supports that comparison; it does not prove what happened in an earlier request with missing context evidence.
Retrieval supplies guidance for generation. Translation reuse retains existing target text, while prompt caching reuses provider input processing. None of those operations trains the model on every correction or turns the localization engine into a general web-ingestion service.