Open-Source Engine Deployment Plan (Layer L2)
Both L2 candidates are free open source and can run offline on this machine. This page is comparison and steps only; every install action awaits boss approval (PENDING).
Argos Translate vs LibreTranslate
| Item | Argos Translate | LibreTranslate |
|---|---|---|
| What it is | A Python translation library that runs offline, installed via pip | A self-hosted translation API service (also powered by Argos models), REST interface |
| Installation | pip install argostranslate + import .argosmodel model packages | pip install libretranslate or Docker; listens on a local port (e.g. :5000) after start |
| Footprint / model size | Program tens of MB; each language-pair model about 100–300 MB (estimate), download on demand | Same (reuses Argos models); Docker image adds about 1–2 GB (estimate) |
| Offline | ✅ Fully offline once models are downloaded | ✅ Fully offline (local service) |
| Quality expectation | Good for short general sentences; Indonesian is medium; industry terms must be protected by the L1 dictionary | Same as Argos (same models); adds an HTTP API so every module can call it |
| Best fit | Backend scripts, batch jobs, queue workers call it directly via import | Many modules share one translation HTTP service (mall/CS/tickets all call it) |
| Language pairs we need | zh↔en and en↔id have ready models; zh↔id has no direct model and goes through English pivot (zh→en→id); the quality loss is patched by L1 dictionary post-processing. | |
Installation Steps (all PENDING)
Download the zh↔en and en↔id .argosmodel packages (about 100–300 MB each, estimate) and archive them in runtime/translate-models/.
pip install argostranslate; then argospm install pointing at the local .argosmodel files, no network needed.
pip install libretranslate, listen on 127.0.0.1 only, never exposed; reuse the same model directory to provide one translation API for mall/CS/ticket modules.
Run 100 real business sentences (quotation/logistics/tax); term accuracy must be 100% (protected by L1); sentence readability is manually sampled and reported to the boss.
Offline package plan
If the server has no internet: download models on an online computer → copy via USB / intranet to the server → install locally; the server never needs to go online. Model packages and install scripts join the backup system.