Download the PHP package tonydev/lara-glot without Composer

LaraGlot

Automatic translation for Laravel. LaraGlot translates Eloquent models with JSON translatable attributes, nested page-builder content, and static PHP language files — and it goes to great lengths to never pay for the same translation twice.

Five drivers ship out of the box: Anthropic Claude, OpenAI, DeepL, Google Translate, and Ollama for local models. Laravel placeholders, URLs, HTML tags, and your brand terms survive translation intact.

Table of Contents

• Requirements
• Installation
• Configuration
• Drivers
• Keeping Costs Down
• Glossary and Brand Terms
• Quality Review Pass
• Placeholder Protection
• Model Setup
• Language File Translation
• Artisan Commands
• Queue Setup
• Filament Integration
• Custom Drivers
• Custom Prompts
• Failure Behaviour
• Upgrading to 2.1
• Upgrading from 1.x
• License

Requirements

Installation

The service provider is auto-discovered. Then publish the config:

Two drivers need an extra package; the rest use Laravel's HTTP client directly:

If you use queued model translation, make sure job batching is migrated:

Running php artisan migrate also creates the lara_glot_translation_retries table, which tracks values that fell back to source text so laraglot:retry can re-attempt them. It's optional — without it, translation still works and LaraGlot logs a single notice instead of tracking fallbacks.

Configuration

Pick a driver and set its key in .env:

Every driver accepts the same tuning knobs in config/lara-glot.php: chunk size, retries, retry delay, concurrency, HTTP timeout, and cache TTL. The defaults are sensible; you rarely need to touch them.

For slow reasoning models (e.g. DeepSeek-class models behind an OpenAI-compatible endpoint), raise the timeout and/or lower the chunk size:

When a chunk still times out (or gets rate-limited), the LLM drivers automatically retry it in smaller sub-chunks before falling back, so one slow request no longer drops a whole chunk of strings to source text.

Fallback detection

When a driver fails on a string, that string falls back to the source text. LaraGlot refuses to let that go unnoticed:

• laraglot:files --sync prints a per-file translated / identical-to-source count, warns when the fallback share exceeds lara-glot.fallback.warn_ratio (default 0.5), and exits non-zero on failures.
• When every translatable string in a file comes back identical to the source — the signature of a failed run — the file is not written and the job fails so it can retry, instead of an untranslated English file masquerading as done. Disable via LARAGLOT_FAIL_ON_FULL_FALLBACK=false if a locale is legitimately identical to the source.

Protected values

Values that are references rather than prose — 'link' => 'route:contact', bare URLs, mailto:/tel: links — are passed through untranslated. Patterns are configurable in lara-glot.protected_value_patterns.

Drivers

To use another OpenAI-compatible provider, change the base URL — no code changes:

Keeping Costs Down

This is where LaraGlot earns its keep. A translation API call is made only when a string has genuinely never been translated before:

1. In-process cache. Within one request or job, a repeated string costs a RAM lookup, nothing more.
2. Persistent cache. Every translation is stored in your Laravel cache (30 days by default), keyed per string, per language pair, per driver. Re-running a command, re-saving a model, or rebuilding language files re-uses everything already translated.
3. In-batch deduplication. If "Save" appears forty times across your language files, it is sent to the API once and fanned back out to all forty keys.
4. Change detection. Model translation tracks a hash of the source content and skips fields that haven't changed since the last run.
5. Per-string caching, not per-chunk. Editing one string in a file invalidates that string only — the other forty-nine stay cached.

In practice a second full run over an unchanged project costs zero API calls. You can verify this yourself: every driver exposes getStats() with cache hits, misses, and the actual API call count.

Glossary and Brand Terms

Two tools to keep terminology under control:

protected_terms are shielded with the same token mechanism used for URLs, so even Google and DeepL can't mangle them. Matching is case-sensitive and whole-word: LaraGlot is protected, laraglots is not.

Changing the glossary automatically invalidates the affected cached translations — you never serve stale wording after a terminology decision.

Quality Review Pass

Machine translation is accurate but often stiff. With the review pass enabled, each freshly translated chunk gets a second call where the model acts as a native-speaker reviewer: it fixes overly literal phrasing, replaces vocabulary no real product would use, and returns already-good translations unchanged.

Cost notes, because they matter:

• The review only runs on strings that weren't cached — and the reviewed result is what gets cached. You pay the extra call once per string, ever.
• If the review call fails for any reason, the first draft is kept. A translation is never lost to the reviewer.
• LLM drivers only; Google and DeepL ignore the flag.

Placeholder Protection

Translation engines will happily turn :count into :nombre and break your app. Before any text leaves your server, LaraGlot replaces sensitive tokens with opaque markers, and swaps them back afterwards:

Input:

Sent to the API:

After restore:

This is a hard guarantee enforced in code, not a prompt instruction, and it applies to every driver and every translation path.

Model Setup

Add the trait to any model using spatie/laravel-translatable:

Register it in config/lara-glot.php:

Nested JSON content (page-builder sections, repeaters) is translated recursively. Keys listed in ignored_keys — identifiers, security values, media paths, locale/format config, geo and ordering fields, timestamps — are passed through untouched. The published config ships with a comprehensive list of keys that are never human-readable labels (id, uuid, slug, sku, token, avatar, latitude, created_at, …); add your own app-specific keys there. Deliberately absent are words like email or price that are often translatable labels in language files.

Language File Translation

Translates your lang/en/*.php files into every configured locale, preserving nested key structure:

Files listed in exclude_files (framework files like validation) are skipped.

Artisan Commands

laraglot:sync

Scans registered models and dispatches translation jobs for anything new or changed.

laraglot:files

Translates PHP language files. See the section above for options (file?, --locale=, --force, --sync).

laraglot:audit

Reports every language-file key and model attribute whose value is byte-identical to the source locale — the signature of a silent translation fallback — and can repair them in place:

• Exits non-zero when any file/model locale is flagged at or above the threshold, so it works as a CI gate or scheduled canary (Schedule::command('laraglot:audit')).
• --repair re-translates only the values still equal to the source (missing keys included) with a forced cache refresh, so fixing fallbacks never re-bills the whole corpus and evicts poisoned cache entries.
• Values that are supposed to be identical across locales belong in lara-glot.audit.allowlist — they're never flagged or re-billed. The default list ships with common acronyms, units, and platform names ("OK", "SMS", "PDF", "iOS", "PayPal", …); extend it in the published config. ignored_keys and protected_value_patterns are excluded automatically.

laraglot:retry

Whenever a value falls back to source text during a translation run, LaraGlot records it in the lara_glot_translation_retries table — a fallback is never silent. This command re-attempts those units:

• Bounded, with back-off: each failed attempt reschedules the unit at base_delay_minutes × 2^attempts (default 30 min base); after max_attempts (default 5) the unit is parked as exhausted and the command exits non-zero so it surfaces for human review instead of re-billing forever. Tune via lara-glot.retry.
• Never wastes an API call: units whose source text changed, whose key was removed, or that were already healed elsewhere (an audit --repair, a --force run, a manual edit) are resolved without touching the driver.
• Idempotent + resumable: run it on a schedule (Schedule::command('laraglot:retry')->hourly()) and coverage converges; exhausted strings that are legitimately identical belong in lara-glot.audit.allowlist.

Queue Setup

Heavy work runs on Laravel queues (translations by default):

For production, run the worker under Supervisor with --timeout above the model job timeout (600s).

Filament Integration

Register the plugin in your panel provider:

You get a management page showing translation coverage per model and language, with actions to queue missing or forced re-translations.

Custom Drivers

Register your own engine without forking the package. In a service provider's boot():

Then set LARAGLOT_DRIVER=my-engine. The driver needs to implement Tonydev\LaraGlot\Contracts\TranslationDriverInterface — two methods, translate() and translateBatch(). Extend AbstractLlmDriver instead if your engine is prompt-based and you want chunking, caching, deduplication, glossary support, and the review pass for free; then the only method to write is sendChunk().

Custom Prompts

The built-in LLM prompt asks for translations that read like a native speaker wrote them, matching the register of the source text. If you want full control:

The string replaces the built-in prompt entirely; {source} and {target} are substituted. Changing it invalidates the affected cache entries, same as glossary changes.

Failure Behaviour

Two rules, depending on where you are:

At render time (a page reading a translation), failure is graceful: a failed translation returns the original string, never an exception and never half-translated token soup. Placeholders are restored even on the fallback path — users never see __VAR_0__.

In jobs and commands, failure is loud — a batch that writes English into lang/ar/ while reporting success is corruption, not resilience:

• Chunks retry with exponential back-off; transient failures (timeouts, 429s, 5xx) additionally retry in halved sub-chunks before any string falls back.
• A batch driver error throws — the job fails visibly and retries, and nothing gets cached or written.
• A file whose output is 100% identical to the source is treated as a failed run: not written, job failed.
• Any individual string that does fall back is recorded in the retry table and re-attempted by laraglot:retry — a fallback is acceptable, a silent fallback is not.
• Corrupt cache entries read as misses and are evicted automatically; forced refreshes (--force, --repair, retries) bypass every cache layer, including the driver's own.
• Everything is logged under the [LaraGlot:Driver] tag.

Upgrading to 2.1

2.1 fixes a data-corrupting bug and tightens failure behaviour. Read this if you ran 2.0 in production:

• Critical fix — multibyte corruption: under Laravel's default process concurrency driver, translations to Arabic, Chinese, Hindi, and other multibyte scripts were silently replaced with the English source (unserialize(): Error at offset … in the log was the only trace). Fixed for both file and model translation. Audit your existing locales — php artisan laraglot:audit reports the damage and --repair heals only the affected values.
• Filament v4/v5 only: the admin page uses the Schemas API and a non-static $view; Filament 3 panels are no longer supported.
• Run php artisan migrate: creates the optional lara_glot_translation_retries table used by laraglot:retry.
• Behaviour change: batch translation failures in jobs/commands now throw instead of silently writing source text; a 100% source-identical file fails the job instead of being written. Escape hatch: LARAGLOT_FAIL_ON_FULL_FALLBACK=false.
• ignored_keys defaults grew substantially (identifiers, media paths, timestamps, …). If any of the new defaults collide with translatable label keys in your app, remove them in your published config.
• New config blocks — retry, audit (threshold + allowlist), fallback, protected_value_patterns, per-driver timeout, drivers.google.max_length — all optional with sensible defaults; re-publish the config to pick them up.

Upgrading from 1.x

• ignored_keys defaults shrank to a generic list. If you relied on removed defaults (cta_url, is_active, and similar app-specific keys), add them to your published config.
• Ollama caching changed from chunk-level to string-level. Old chunk cache entries are simply ignored; strings re-cache individually on the next run.
• New config blocks (glossary, review, prompt, drivers.anthropic) are optional — re-publish the config or copy them in if you want the new features.
• Model defaults moved to current generations (gpt-5-mini, llama3.2). Pin your previous model via env if you need to stay put.

License

MIT. Created by TONY THE DEVELOPER