Skip to content

v1.10.361 – AI term extraction (bilingual, reviewable), one project-termbase truth, batch-aware AutoPrompt

Choose a tag to compare

View all tags
@michaelbeijer michaelbeijer released this 20 Jul 23:12
v1.10.361
0db0eb0

This release rebuilds term extraction around your AI provider, unifies the app's idea of what the project termbase is, and fixes a long-standing defect in AutoPrompt-generated prompts (v1.10.352 → v1.10.361).

AI term extraction – a project glossary in one click

🔍 Extract Terms now asks your configured AI model for a project glossary instead of running the old mechanical frequency analysis:

  • Bilingual results – the model returns source → target term pairs (actual translations), so the created termbase starts producing TermLens matches immediately. Previously every extracted entry had an empty target and matched nothing.
  • Review before committing – extracted pairs land in an editable table: fix entries in place, tick/untick rows (now with bulk tick/untick), then Create Project Termbase saves exactly what you see. Where the model is unsure, it leaves the target empty with a note rather than guessing.
  • Language pair prefilled from the project, optional Domain / Subject hint, background AI call (no frozen dialogue), clear error reporting, and a visible warning if a very large project is truncated for analysis.
  • On the way there, the whole extraction flow was unbroken: the permanently greyed-out button (v1.10.352), the dead in-dialogue Extract Terms button (v1.10.353), the ignored "Project-specific" scope on + Create New (v1.10.354), capitalised terms never being recognised (v1.10.355), and Create Project Termbase doing nothing (v1.10.356).
  • The mechanical extractor is retired – prefer fully-local extraction? Point the AI provider at a local model (Ollama).

One definition of "project termbase"

The app can no longer disagree with itself about which termbase is the project termbase. The pink Project tick in the Termbases tab is now the single authoritative definition everywhere – creation guards, term extraction, matching, ranking, and the entry editor all read it, and every way of assigning the role goes through one manager method. Stale legacy flags (which could block term extraction with a phantom "project termbase already exists") are repaired at startup, and termbases with Read unticked no longer sneak into matching. Locked in by new regression tests.

AutoPrompt: batch-aware prompts

AutoPrompt-generated prompts wrongly told the translator AI it receives text "one segment at a time, in isolation" – but Batch Translate/Proofread actually send numbered batches of segments. Generated prompts therefore forbade using context the AI could legitimately see, and left terminology "choices" open that can't stay consistent across batches. The template now describes batched delivery correctly: in-batch context may be used, document-wide checks belong to a QA pass, every recurring term must be locked (no open "X or Y" choices), and ⟦TC: …⟧ correction markers stay attached to their own segment. Existing generated prompts in your library keep the old wording – regenerate the ones you rely on. (Same fix shipped in Supervertaler for Trados v18/19.20.109 – the two products share this template.)

Install

  • Windows standalone: download the .zip below, extract, run Supervertaler\Supervertaler.exe (do not move the EXE out of its folder).
  • macOS standalone: download the .dmg below.

Full changelog: https://github.com/Supervertaler/Supervertaler-Workbench/blob/main/CHANGELOG.md

Assets 3
Loading