# Which Content Leaves Your Computer, and What Doesn’t
When you paste text or drag a file into a modern translation interface, the immediate assumption is that the content travels somewhere. For professionals handling sensitive drafts, proprietary terminology, or regulated documents, that uncertainty creates friction. This guide explains exactly how JoesanTrans handles your data, why the pipeline is structured the way it is, and what you should watch for when switching between workflows.
The Five-Layer Translation Pipeline
Every translation request follows a fixed cascade before returning a result. The system does not guess which engine might be fastest; instead, it checks five distinct sources in order:
- Local Cache
- Offline Dictionary
- Translation Memory
- Offline Model
- AI Layer
This sequence exists to maximize speed and minimize dependency on an active connection. When you type a phrase, the software first scans the local cache. If a recent translation for that exact string already exists in memory, it returns instantly. If not, it queries the offline dictionary next. Dictionaries handle individual words, compound terms, and short phrases. Once words pass through those two gates, the system consults your personal translation memory. Finally, it attempts a full sentence-level parse using the downloaded offline model. Only if every local step fails does the request reach the AI layer.
You can track exactly which stage resolved each output. Every translated line displays a small indicator showing its origin. When you see a cache icon, you know the system retrieved a previous result without any network activity. A dictionary badge means a lookup occurred offline. A memory tag confirms your own edited version was reused. An offline model marker indicates a neural inference ran entirely on your hardware. The AI badge appears only when the request finally left your machine. Reading these markers helps you predict latency and understand how your settings affect performance.
What Stays Strictly Local
The first four layers operate independently of any network. You do not need Wi-Fi, cellular service, or a proxy to run them. They also do not consume any cloud quota. This architectural choice was made because most professional workloads involve repetitive phrasing, industry-specific jargon, or internal style guidelines that never change mid-session. Keeping these components local removes billing surprises and guarantees consistency even in restricted environments.
Your offline dictionary covers seventy-two translation directions. It includes regional variants such as Cantonese and Simplified Chinese terms alongside traditional counterparts. The dictionary updates automatically when you install new language packs, but it never phones home to report which entries you searched. It simply answers lookups against the files stored in your application folder.
Your translation memory works differently from automated crowdsourcing platforms. It does not scrape public texts or aggregate "high-quality" examples from unknown authors. Instead, it records only the exact strings you manually correct after a translation appears. If you adjust a generated sentence and save it, the system binds that corrected version to the original source. The next time you encounter the same source text, the memory layer intercepts the request and serves your preferred phrasing directly. Over time, this trains the software to reflect your actual working habits rather than generalized statistical averages.
Offline models provide the heavy lifting for sentence reconstruction. Depending on your target language pair, a single-direction model ranges from thirty-seven to seventy-nine megabytes. If you frequently switch between multiple languages during the same project, a cross-directional universal model spans approximately four hundred ninety-seven megabytes. Both options are downloadable on demand. You choose which languages to install, and the software reserves disk space accordingly. No background indexing occurs until you explicitly trigger the download.
When Data Travels to the Cloud
Text leaves your computer exclusively when it reaches the fifth layer. The AI component handles complex syntactic restructuring, ambiguous polysemes, or highly domain-specific passages that fall outside dictionary coverage, memory matches, and offline model training bounds. Because cloud inference requires authentication, the application delegates all key management to secure servers. Your desktop client never stores API credentials, session tokens, or authorization headers. The moment a translation completes, the temporary routing information clears automatically.
Consumption patterns depend entirely on your account tier. The free tier grants unrestricted access to all offline capabilities. You may download every available dictionary, build a complete translation memory, and load multiple offline models without restriction. The only limitation applies to the AI layer, which operates under a daily character allowance. Once that threshold is reached, the pipeline still cascades through the first four stages, but any remaining unresolved segments remain untranslated until the quota resets or you upgrade.
VIP subscribers receive expanded parameters for both cloud and advanced desktop features. Offline models become unlimited, meaning you can store dozens of direction-specific weights without choosing between them. Translation memory and custom glossaries lift their caps, allowing enterprise teams to maintain sprawling terminology databases. Document translation and screenshot recognition also remove their frequency restrictions, enabling bulk processing without manual pagination.
Understanding this boundary prevents unnecessary anxiety about privacy while keeping budget expectations realistic. If you primarily localize technical manuals with consistent phrasing, the cache and memory layers will resolve the vast majority of content without ever dialing out. If you draft marketing copy with high lexical variability, you will rely more heavily on the AI tier and should monitor your daily character consumption accordingly.
Navigating the Interface and Workflow Habits
JoesanTrans runs natively on macOS and Windows, with a unified codebase that ensures consistent behavior across operating systems. The graphical layer supports ten interface languages, allowing non-English speakers to configure menus, dialogues, and error messages in their preferred locale without altering the underlying translation pipeline. Regardless of UI language, the origin indicators function identically, and storage paths follow native filesystem conventions.
Practical workflow adjustments stem from recognizing how each layer interacts with your input habits. When you notice frequent AI badges for predictable sentence structures, review your translation memory entries. Manual corrections carry more weight than automatic suggestions because the memory layer prioritizes explicit user edits. Deleting outdated memory items and rewriting problematic phrases forces the cascade to rebuild its local knowledge base rather than relying on stale outputs.
Managing offline models benefits from a deliberate installation strategy. If your projects concentrate on a single language pair, downloading the smaller single-direction files conserves drive space and reduces parsing overhead. Universal models shine when context switches occur rapidly, though the initial forty-nine-megabyte footprint demands adequate free storage. The application warns you before commencing large transfers, preventing interrupted downloads from corrupting cached inference weights.
Tracking character usage near your daily cap requires checking the status panel before initiating bulk operations. If you anticipate high-volume AI reliance, schedule large exports during reset windows or evaluate whether upgrading unlocks the broader document translation privileges that streamline batch processing. Conversely, users who prefer strict isolation can disable automatic network prompts and rely entirely on cached lookups, accepting slightly longer resolution times in exchange for guaranteed local execution.
Control over your data boundary comes from observing the pipeline, adjusting local resources, and aligning tier selection with actual traffic patterns. The software gives you clear signals at every step. Follow those signals, refine your manual corrections, and you will consistently route routine content through the fastest, safest lanes while reserving cloud inference for the cases that truly require it.