Autonomous Translation Infrastructure

Every translation is cached the moment it's generated — by source text and target language — so the same phrase is never sent to an AI model twice.

``` First translation of "Free shipping over $50" into Spanish: → Not in the cache → AI translates it → cached for next time

Second translation of "Free shipping over $50" into Spanish: → Found in the cache → reused instantly → $0 AI cost, reuse count increments

AI cost math: Without Translation Memory: 10,000 products × 5 fields × 10 languages = 500,000 AI calls $0.002/call = $1,000/month

With Translation Memory (after first pass): 500,000 cache lookups = $0 AI calls only for NEW unique text If 20% unique per month: 100,000 × $0.002 = $200/month

Savings: $800/month = $9,600/year ```

Every cached entry tracks how many times it's been reused, when it was last used, and its review status — so you can see exactly how much AI spend Translation Memory is saving over time.

Every glossary term is one of three rule types:

• `do_not_translate` — the term is kept exactly as written in every language.
• `always_translate_as` — the term maps to a specific translation per language you define.
• `guidance` — free-text instructions the AI follows whenever the term appears (e.g., register, tone, or context-specific meaning).

These rules are injected directly into the AI's translation prompt — not corrected afterward — so they're enforced from the first draft, not applied as a post-processing patch.

Example in action:

``` Source: "Aeion OS is the best SaaS Platform for Enterprise" Glossary: "Aeion OS" → do_not_translate "SaaS Platform" → always_translate_as "Enterprise" → guidance ("use 'Großunternehmen' for companies with 500+ employees")

↓ AI translation (es/de blend for illustration): "Aeion OS es la mejor Plataforma SaaS para Großunternehmen" ✅ "Aeion OS" kept exactly (do_not_translate) ✅ "SaaS Platform" → "Plataforma SaaS" (always_translate_as per your language map) ✅ "Enterprise" → "Großunternehmen" (guidance applied for company size)

Without a glossary, generic AI translation drifts: "Aeion OS es la mejor plataforma SaaS para empresas" ❌ inconsistent with your brand voice ```

Managing your glossary:

``` Admin → Localization → Glossary

Import CSV: term, ruleType, targetTranslations, description, matchCase "Your Brand", do_not_translate, , , true "SaaS Platform", always_translate_as, {"es":"Plataforma SaaS","de":"SaaS-Plattform"}, , true "Enterprise", guidance, , "Use the formal register for companies >500 employees", false

Bulk export for translator review Term usage analytics ```

Whenever a localized field is saved without its target-language translation, Aeion notices and queues the translation automatically — no manual trigger required.

What happens when you save a page:

``` 1. You save a page in English with localized fields: title: { en: "Free Shipping", es: null, fr: null }

1. Aeion detects:

1. Translation jobs are queued (non-blocking — your save completes instantly):

1. Workers process the jobs in the background:

1. Done in ~30 seconds. You didn't lift a finger.

Translation jobs run on a background queue with several workers processing in parallel, so saving content is never blocked waiting on AI calls.

Built-in duplicate-prevention:

• Save the same content twice within a short window → only one translation job runs, not two.
• Resubmit an identical batch import → it's recognized and not duplicated.

Failed jobs retry automatically with increasing delays between attempts, and every job — active, completed, or failed — stays visible in your admin queue view.

Why parallel workers instead of one at a time:

``` AI provider rate limits: typically tens to hundreds of requests/minute depending on provider Content volume: most tenants save a handful of items per minute → A handful of parallel workers keeps pace with normal usage → Plenty of headroom for bursts (bulk imports, catalog launches)

High-volume tenants can add dedicated workers so they aren't sharing capacity with anyone else. ```