CatScribe Docs

#Performance Guidelines

Translation speed depends on document size, engine choice, AI refinement settings, and your computer. This page lists the app's real limits — the enforced hard caps and the tunable settings — plus practical advice for large projects.

#Hard Limits (Enforced, Not Suggestions)

These are hard caps. Files or text over these limits are rejected, not processed slowly. The in-app summary reads: "Documents up to 50MB · Subtitles up to 5MB · Videos up to 500MB."

Limit Value
Document upload (File Translation: PDF, DOCX, EPUB) 50 MB
Subtitle file upload (SRT, VTT) 5 MB
Video file upload (Subtitle Translation) 500 MB
Quick Translate source text 8,192 characters ("Text exceeds the maximum length of 8,192 characters.")
Chunks per job 500 budgeted; 1000 hard limit (large texts are merged into bigger chunks, never dropped)
Chunk size setting Backend clamps to 1000–10000 characters

PDFs additionally have complexity guards. A PDF that exceeds them fails fast with "This PDF is too large or complex to translate in a single pass." and the suggested fix is exactly what works: split the PDF into smaller files (fewer pages) and translate the parts. A structurally broken file reports "This PDF appears to be damaged or invalid."

If a document exceeds the 50 MB cap, split it before uploading — see Working With EPUB, DOCX, and PDF.

#Hardware Expectations

The Providers/AI Models screen shows "Your Hardware Profile" and classifies your machine:

Tier Criteria Practical meaning
"High-End Hardware" GPU + 16 GB+ RAM All engines and large local LLMs are practical; Maximum quality is realistic for full books
"Mid-Range Hardware" 12 GB+ RAM All MT engines fine; mid-size LLMs (~7–8B) for AI layers
"Low-End Hardware" Below that Prefer Argos and Fast/Balanced; use small LLMs (~1–3B) if any

RAM is the binding constraint for local engines:

  • Argos is the lightest engine and the best default on modest hardware.
  • MarianMT/NLLB load neural models into system RAM (roughly 1–2 GB while active) and run on CPU; the app shows a "High RAM Usage" warning when you select them.
  • Ollama/LM Studio models need RAM (or VRAM) proportional to model size — a 7B model wants ~5 GB, a 70B model ~40 GB. Follow the "Recommended" badges, which are matched to your tier.
  • Disk space matters too: language packs (50–100 MB each), Marian/NLLB weights (0.2–1.2 GB), COMET (~1–2 GB), and Ollama models (up to tens of GB) all accumulate — the AI Models screen tracks "Models Disk Usage".

#What Slows Jobs Down Most

Roughly in order of impact:

  1. AI refinement layers — every enabled stage (validation, rewrite, technical) multiplies per-chunk work. Maximum quality can be several times slower than Fast.
  2. Large models on limited hardware — an LLM that barely fits in RAM thrashes; a smaller model is often faster and better in practice.
  3. Scanned or complex PDFs — OCR and high-fidelity layout modes cost far more than clean text extraction. EPUB and DOCX are cheaper formats than PDF for the same content.
  4. Oversized chunks — approaching model context limits causes slow responses and retries.
  5. Background load — other heavy apps competing for the same RAM/CPU.

Retries also cost time by design: failed chunks back off ~5 s/15 s/45 s, and rate-limited providers can pause up to 5 minutes. A job that keeps retrying is telling you something — check the provider's status instead of waiting it out; see Troubleshooting.

#Tips For Large Projects

  • Test small first. Run a chapter-sized sample and check quality, formatting, and speed before committing to the full book.
  • Use Fast or Balanced for the first draft, then improve only the chunks that need it with Improve With AI or Chunk Overrides — much cheaper than re-running everything on Maximum.
  • Don't re-run the whole job after glossary changes — re-run affected chunks.
  • Enable "Prevent system sleep while translating" (Settings) for long unattended runs; "Auto-Resume Pending Translations" (default on) recovers jobs after restarts.
  • Pause instead of cancelling. Pause/Resume on the Translations screen keeps completed chunks; you can even adjust settings while paused.
  • Follow the "Recommended setup" hint in Auto Mode — it already accounts for your hardware tier.

#Large PDF Expectations

PDF is a visual layout format, not a clean text structure, so it is the slowest and least predictable input. For large PDFs:

  1. Let "PDF Translation Mode" stay on "Auto (recommended)" unless you have a reason to override; "Simple text" is fastest for plain-text PDFs, and "Fast (OCR for scanned PDFs)" handles scans.
  2. Translate a small sample and check paragraphs, headings, and tables before the full run.
  3. If the file trips the "too large or complex" guard, split it by page ranges and translate the parts.
  4. When you have the same content in EPUB or DOCX, prefer that source — see Working With EPUB, DOCX, and PDF.