Technical specifications and troubleshooting – Tracal
← Back to Tracal

Technical specifications and troubleshooting

This page explains Tracal's translation inference engine, the Shared Model architecture and its settings.

1. Technical specifications and architecture

Translation inference core

  • Inference engine: C++ native llama.rn high-speed engine.
  • Supported models: TranslateGemma 4B, TranslateGemma 4B (High Precision), TranslateGemma 12B.
  • Memory optimization: Kept stable in memory with use_mlock.

Shared Model specifications

  • Shared identifier: com.localai.sharedmodels is used.
  • Automatic detection: Models saved by Rocal, Lingcal and others are recognized and usable right away.
  • Storage savings: Multi-gigabyte models are shared without downloading them twice.

Translation controls and settings

  • Glossary: Two-way mapping of proper nouns and technical terms (Premium only).
  • Tone: Real-time revision into polite business style, academic style or casual speech.
  • Speaker gender: Pronouns and sentence endings are automatically adjusted to suit the context and gender.

Data processing boundaries (three security layers)

To keep translated text and confidential data confidential, Tracal clearly defines where data is stored and processed.

Layer 1. Communication outside the device (limited)

Processing that accesses outside servers.*There is no processing at all that sends the text or documents you enter, before or after translation, outside.

  • No ad or tracking communication: This app contains no third-party ad or analytics SDKs.
  • Downloading AI models: The app communicates with model servers only when downloading a translation model (such as TranslateGemma) for the first time.

Layer 2. Access within the device, outside the app

Processing that involves no network communication but uses OS file areas and shared storage, or calls native APIs.

  • Model sharing between series apps (Shared Model): Large translation models are shared on the device with the other series apps (Rocal/Lingcal) through shared storage (com.localai.sharedmodels).
  • Importing documents and text: Contracts and text files to be translated are safely brought into the app.
  • Clipboard: You copy translation results to the clipboard yourself.

Layer 3. In-app sandbox (fully private)

Secure processing that is completed only within Tracal's own encrypted local storage and app memory, without even involving OS features.

  • Running translation inference: From parsing the source text to generating the translation, all computation is completed only within the device's CPU and RAM.
  • Storing the two-way glossary: Registered translations of technical and in-house terms are kept in the app's own protected database.

2. Troubleshooting & tips

A guide to adjusting translation accuracy, refining the style prompt, and setting gender and the glossary.

Model ChoiceTranslations aren't accurate or expressive enough

Cause

The AI model currently loaded may be small and unable to keep track of long, complex text or context.

What to do

Switch to a larger model (for example, TranslateGemma 12B) in Model Management.
*A larger model greatly improves accuracy and naturalness, but uses more of the device's RAM and changes the inference speed.

AI RefinementI want to change the nuance or wording of a translation

Cause

The standard literal translation prompt may not reflect the style needed for the purpose (polite business style, academic style, casual speech and so on).

What to do

Use "AI instructions (Refine)" below the translation. Type an extra instruction such as "Make it more polite for business", "Summarize it briefly" or "Make it more casual", and the AI immediately revises and fine-tunes the translation.

GlossaryTechnical terms or in-house names are mistranslated or inconsistent

Cause

Product names and industry terms that aren't in the LLM's general training data get translated literally by ordinary language rules.

What to do

Open the two-way Glossary (*Premium only) and register pairs of source terms and the translations you want. When translating, the AI automatically follows the glossary, preventing inconsistent spellings and mistranslations.

Gender & ToneI want gendered expressions and politeness to match the context

Cause

In some languages, gendered and honorific forms of first- and second-person words and sentence endings change with the context, the speaker and the listener.

What to do

Switch the Gender Selector and Tone settings on the screen. Natural gendered and honorific forms that fit the context and purpose are applied automatically.

3. Frequently asked questions (FAQ)

Common questions and answers about using Tracal.

Q Is the text or document I translate ever sent outside?
No, never. Tracal runs inference 100% on your device, so not a single byte of your source text, the generated translation or your glossary data is sent to outside servers.
Q How do the free (Lite) and paid (Premium) versions differ in security and supported models?
The free version (Lite) has the same security design: fully local processing with zero communication. The paid version (Premium) unlocks translating long documents in one go, custom tones (your own styles) and automatic glossary application.
Q Can I share AI models downloaded in other series apps (such as Rocal)?
Yes. Because Tracal uses the `Shared Model` identifier (`com.localai.sharedmodels`), .gguf model files saved by Rocal or Lingcal are detected automatically and shared, without downloading them twice.
Q Do I ever need an internet connection to translate?
Only when downloading an AI translation model (such as TranslateGemma) for the first time and when updating it. Once the model is saved, every translation feature works in airplane mode or fully offline.