#Improve With AI

CatScribe uses local or connected AI language models in two places: as optional refinement layers that run during a translation job, and as a per-chunk "Improve with AI" action inside the CAT Editor. Both are editorial assistants — the reviewer remains responsible for approving meaning, terminology, and final quality.

#The Per-Chunk Action In The CAT Editor

The "Improve with AI" button sits in the CAT Editor toolbar and works on the current chunk:

1. Click "Improve with AI". A modal opens with a model selector ("Select a model" — Ollama, LM Studio, and Claude models appear together) and an "Optional instruction" field.
2. Optionally type an instruction — the placeholder gives the idea: "Example: make the text more natural but keep names unchanged." The help text explains: "Your instruction is added to the AI enhancement prompt."
3. Click "Improve text". The button shows "Improving..." while it runs (up to 3 minutes for slow local models).
4. The result replaces the chunk's target text and is saved as a new entry in Version History with the note "Improved with AI" — so you can always Revert.

What happens under the hood: the chunk runs through a two-stage validation + rewrite pipeline with glossary verification enabled, so attached glossary terms are respected and re-checked. Your optional instruction is appended to the prompt.

Possible outcomes besides success:

• "AI did not suggest any changes to the text." — the model considered the text fine as-is.
• "AI did not return improved text." / "Failed to improve text with AI." — the model returned nothing usable; try another model or a clearer instruction.

Use the "Layers" button afterwards to open the "Translation Layers Inspector" — it shows a "First Layer vs Improve With AI" diff and the chunk's full modification timeline.

#The Job-Level Refinement Layers

Independently of the editor, File Translation can run AI refinement stages over every chunk during the job:

• With Auto Mode ON: Balanced runs Validation + Rewrite, Maximum runs all three, Fast runs none.
• With Auto Mode OFF in Advanced interface mode you configure stages and per-stage models manually, including Orchestration: "Single-model" ("One resident model handles validation, rewrite, and technical checks with isolated prompts.") or "Multi-model" ("Use separate per-stage models for advanced validation, rewrite, and technical review.").
• In Simple interface mode, enabling refinement auto-configures everything: CatScribe picks the preferred model and skips stages when no issues are found — per-stage pickers are never shown.
• The "Smart Pipeline Quality Threshold" setting (default 85, range 70–95) lets the pipeline skip the expensive rewrite stage when a chunk's quality score is already high.

#Where AI Improvement Helps Most

• Dialogue that sounds stiff or unnatural.
• Long sentences that need clearer target-language flow.
• Segments where the meaning is right but the tone is wrong.
• Draft translations that need a readability pass before approval.

#Where It Is Not The Right Tool

Do not reach for AI improvement first when the problem is:

• Missing spaces or merged words caused by source extraction — see Common Issues & Fixes.
• Broken tags, placeholders, or numbers — check the chunk's QA Flags instead.
• Names or terms that keep drifting — that is a Glossaries job.
• A chunk where the base translation failed outright — use Chunk Overrides to re-run it with a different engine.
• Content requiring expert legal, medical, or safety review.

#Safe Revision Pattern

1. Fix glossary and formatting issues first.
2. Run "Improve with AI" on the specific chunk that needs help, with a targeted instruction.
3. Compare the result against the source (and the Layers diff).
4. Keep, edit, or Revert via Version History.
5. Re-check the chunk's score with "Recalculate score" if you want a fresh estimate.
6. Approve and move on.