Changelog

Copy Markdown View Source

All notable changes to this project will be documented in this file.

[0.9.3] - 2026-08-17

Self-hosted vLLM — and any OpenAI-compatible gateway — is now supported end to end. The provider can be configured inline in your application config, and that same provider can still be overridden per session from the dashboard. Since 0.9.2 the LiveDashboard Override LLM Provider form offers a vLLM adapter you can point at a gateway by hand; this release closes the loop, so an instance that already has a gateway in its config no longer has to retype it: the form opens on the configured provider, and anything you do type overrides it.

Added

  • :default_provider — configure the provider inline, override it when you want. An instance whose endpoint and API key already come from its own environment should not make an operator retype them into a dashboard form. Naming the provider your config is for:

    config :gettext_translator, GettextTranslator,
  endpoint: MyApp.LLM.ChatVLLM,
  endpoint_model: System.get_env("GETTEXT_TRANSLATOR_MODEL", "gemma-4-26b-a4b-it"),
  default_provider: :vllm

    opens the Override LLM Provider form with that adapter selected and :endpoint_model filled in, marks API Key and Endpoint URL optional, and treats a blank field as "use what the instance is configured with". Submitting the form unchanged translates against your gateway.

    Crucially, the override then routes through your own :endpoint module rather than the generic LangChain.ChatModels.ChatOpenAI. For a self-hosted gateway that is the difference between working and not: a module like MyApp.LLM.ChatVLLM resolves its endpoint and bearer token itself, whereas ChatOpenAI requires the full /v1/chat/completions URL to be supplied by hand.

    Anything typed into the form still wins, selecting a different provider behaves exactly as before — and never inherits the configured provider's credentials — and leaving :default_provider unset preserves the previous behavior exactly: nothing at all is inherited unless the option opts in.

    The value is validated against Parser.known_providers/0; an unrecognized slug is refused with a logged warning rather than accepted, because it would match no <option> and leave the browser selecting OpenAI while the model stayed prefilled from config.

    Exposed as GettextTranslator.Util.Parser.instance_default/0, provider_label/1, known_providers/0 and resolve_override/2 — the override form's provider/credential resolution moved out of the LiveDashboard page into Parser, where it is config resolution rather than view logic and is directly unit-tested.

Changed

  • Dependencies: langchain 0.9.2 → 0.10.0, phoenix_live_dashboard 0.8.7 → 0.9.0, phoenix_live_view 1.2.6 → 1.2.9, plus transitive phoenix 1.8.11, req 0.7.2, ecto 3.14.2, mint 1.9.3 and plug_crypto 2.2.0.

    The requirements in mix.exs are unchanged and stay deliberately open — langchain >= 0.8.0, phoenix_live_dashboard >= 0.6.0, phoenix_live_view >= 0.17.0 — so host applications remain free to resolve their own versions. Only mix.lock moved.

    LangChain 0.10.0's breaking changes are confined to Message.status — stream failures now report :stream_error rather than :cancelled, and a provider content filter (or an Anthropic "refusal" stop reason) now reports :content_filtered rather than :complete. None of them reach this library: it never enables streaming, registers no callbacks, and never matches on Message.status — only on {:ok, %{last_message: %Message{content: content}}} and {:error, _, reason}.

    The call surface is otherwise untouched. LLMChain.new!/1, add_messages/2 and run/1, Message.new_system!/1 and new_user!/1, ContentPart.parts_to_string/1, and the ChatOpenAI, ChatAnthropic, ChatOllamaAI and ChatGoogleAI adapters all keep their signatures — no public function was added or removed in any module this library calls, and run/1 still returns {:ok, chain} / {:error, chain, error}.

Fixed

  • An unset token env var now reports "token not configured" instead of failing obscurely. Both the GitHub and GitLab pull-request clients read their token with Map.get(config, :key, ""), whose default only applies to a missing key. The usual config shape — github_token: System.get_env("GITHUB_TOKEN") with the variable unset — produces a key that is present and nil, which slipped past the token == "" guard and was sent as a nil HTTP header value, failing deep inside the client. Both clients now coalesce nil to "" and return the intended error.

[0.9.2] - 2026-07-13

Added

  • First-class vLLM provider. The LiveDashboard Override LLM Provider form now offers a vLLM (OpenAI-compatible) adapter alongside OpenAI, Anthropic, Ollama, and Google AI. It maps to LangChain.ChatModels.ChatOpenAI (vLLM speaks the OpenAI wire format) with the endpoint URL you supply, so a self-hosted gateway behind nginx + a bearer token can be selected per session without editing config files. See the new "Using vLLM" section in the README.

Fixed

  • Custom OpenAI-compatible base URLs are now actually used. ChatOpenAI reads its base URL only from the struct passed to new!/1 — it does not resolve one from application env — but the translation chain never threaded the configured endpoint into that struct. As a result, a configured "openai_endpoint" (or a dashboard-override endpoint URL) was silently ignored and every request went to https://api.openai.com/v1/chat/completions. The endpoint is now threaded into the adapter via the new GettextTranslator.Processor.LLM.build_llm_attrs/1, honoring both the "openai_endpoint" (documented config) and "endpoint" (dashboard override) keys. This makes the documented custom-endpoint configuration — and the new vLLM option — work as described.

[0.9.1] - 2026-07-11

Fixed

  • Higher plural forms are no longer corrupted on dashboard save. Saving from the LiveDashboard rewrites every already-translated entry, and for plural messages the third-and-up forms (msgstr[2]+) were rebuilt from the store — which only tracks msgstr[0] and msgstr[1] — so msgstr[2] was silently overwritten with a copy of msgstr[1] (e.g. Ukrainian "кредитів" became "кредити"). PoHelper.update_po_message/2 now preserves any non-empty higher plural forms already present in the PO file and only fills them with the generic plural translation when they are still empty. The hardcoded uk/ru/pl language list is gone: the PO file's own msgstr[n] keys (from Plural-Forms/nplurals) define which forms exist, so all languages — including 6-form ones like Arabic — are handled correctly. An explicit :plural_translation_2 on a translation entry still overrides msgstr[2] when provided.
  • mix gettext_translator.run no longer drops higher plural forms. The CLI processor replaced the whole msgstr map with only forms 0 and 1 when translating pending plural messages, deleting the msgstr[2] line entirely for three-form languages. It now reuses the same form-preserving update logic as the dashboard.
  • Changelog JSON files no longer churn in diffs. Entries whose text and status are unchanged keep their original last_updated timestamp (previously every pending entry was rewritten with a fresh timestamp on each dashboard session), and the JSON output is now deterministically sorted by message id — so translation_changelog/*.json diffs only show real changes instead of hundreds of rewritten lines.

Changed

  • Removed the unreachable [TRANSLATION_FAILED] fallback in the CLI plural path (LLM.translate/2 already falls back to an empty translation on provider errors), fixing an Elixir 1.20 type-inference warning. Cleaned up remaining unused require Logger warnings.
  • jason is now a declared dependency (~> 1.4). It was already used at runtime for changelog persistence but was only available transitively.

Dependencies

  • Verified compatibility with LangChain 0.9.x and bumped the lockfile to 0.9.2. LangChain 0.9.0's breaking changes are confined to telemetry event names and custom ChatModel implementations; the LLMChain.run/2 / message-construction APIs used by this library are unchanged.
  • Lockfile bumps with security fixes for transitive deps: req 0.5.18 → 0.6.2 and mint 1.8.0 → 1.9.1 (HIGH CVEs), plug 1.19.2 → 1.20.3, hpax 1.0.3 → 1.0.4, phoenix 1.8.7 → 1.8.9, finch 0.22.0 → 0.23.0, phoenix_live_view 1.1.31 → 1.2.6, credo 1.7.18 → 1.7.19.

[0.9.0] - 2026-06-01

Changed

  • LangChain requirement widened to >= 0.8.0 (was ~> 0.6.2). The open upper bound means GettextTranslator no longer pins a maximum LangChain version, so host applications are free to upgrade LangChain (0.8.x, 0.9.x, …) without this library blocking dependency resolution. Verified compiling and testing clean against LangChain 0.8.11. The existing ContentPart / LLMChain.run/2 handling already covers the upstream content-structure and return-value changes introduced in 0.4.0.

Added

  • Configurable request retries — New optional :endpoint_retry_count config key, forwarded to the chat model constructor. Since LangChain 0.7.0 disabled Req-level HTTP retries by default and surfaces transient 429/503 errors after retry_count attempts (upstream default: 2), this lets you raise or disable retries for rate-limited endpoints. When omitted, the chat model's own default is used.

Dependencies

  • Bumped tooling/optional dependencies to latest within existing requirements: credo 1.7.16 → 1.7.18, ex_doc 0.40.1 → 0.40.3, phoenix_live_view 1.1.22 → 1.1.31.

[0.8.0] - 2026-03-26

Added

  • TranslateGemma Support — Automatic detection and prompt formatting for Google's TranslateGemma open translation models (4B, 12B, 27B). When the model name contains "translategemma" (case-insensitive), the library uses the model's required prompt structure: a single user message with professional translator persona, source/target language names and ISO codes, and two blank lines before the text to translate. No system message is sent.
  • Source Language Configuration — New :source_language config option to specify the language of msgid strings (defaults to "en"). Used by TranslateGemma to build accurate source-to-target translation prompts.
  • Language Names Utility — New GettextTranslator.Util.LanguageNames module mapping 90+ POSIX locale codes to full language names and ISO/BCP47 codes. Handles regional variants (e.g., pt_BR -> "Portuguese" / pt-BR, zh_CN -> "Chinese (Simplified)" / zh-Hans).

Changed

  • GettextTranslator.Processor.LLM now branches prompt construction based on model type: TranslateGemma models use the dedicated format, all other models use the existing system+user message format unchanged.
  • Provider type updated to include source_language field across parser, translator type, and dashboard resolver.

[0.7.0] - 2026-02-28

Breaking Changes

LangChain Upgrade: 0.5.x → 0.6.0

This release upgrades the LangChain dependency from ~> 0.5.0 to ~> 0.6.0, which includes new features and breaking changes upstream.

Upstream Breaking Changes (LangChain 0.6.0)

  • DeepResearch Source Migration — Inline Source struct replaced with unified Citation/CitationSource structs
  • Perplexity Response Format — Content now returns as ContentPart structs instead of plain strings
  • Google AI Grounding Metadata — Raw metadata moved to message.metadata["grounding_metadata"]

Note: These upstream changes do not affect the core translation workflow of this library, as it does not use DeepResearch, Perplexity, or Google AI grounding features directly.

New Upstream Features Available

  • Citations Support — Unified citation abstraction across all providers (Anthropic, OpenAI, Google AI, Perplexity)
  • Customizable Run Modes — Composable execution modes (WhileNeedsResponse, UntilSuccess, Step, UntilToolUsed)
  • Streaming Reasoning Callbacks — on_llm_reasoning_delta for Azure OpenAI reasoning summaries

Added

  • Quick Prompt Buttons — Locale-aware translation action buttons (Regenerate, Short Version, Rephrase, Synonyms) in the translation details panel
  • PromptTemplates Module — New GettextTranslator.Dashboard.PromptTemplates module with per-locale prompt templates for 7 languages (English, Ukrainian, Spanish, German, Polish, French, Portuguese)
  • Quick Translate Event — New llm_quick_translate LiveView event handler for executing prompt-based translation actions

Changed

  • Renamed "LLM Translate" button to "GPT" for brevity
  • Updated LangChain dependency from ~> 0.5.0 to ~> 0.6.0

[0.6.0] - 2026-02-11

Added

  • Dashboard Tab Navigation — Page-level tabs separating Translation Stats, New Extracted, and New Translated views for a streamlined workflow
  • Extract & Merge from Dashboard — Run mix gettext.extract --merge --no-fuzzy directly from the dashboard UI; automatically switches to the New Extracted tab to show newly discovered strings
  • Batch Translate All Pending — Translate every pending entry in a single click with real-time progress bar tracking; uses sequential processing to respect LLM API rate limits
  • New Translated Review Tab — After batch translation, a dedicated tab appears showing all translations from the current session grouped by language and domain, allowing review, editing, and approval before saving
  • LLM Provider Override — Session-scoped form to switch AI adapter (OpenAI, Anthropic, Ollama, Google AI), model, API key, and endpoint URL without modifying config files
  • Extractor Module — New GettextTranslator.Util.Extractor for dev/prod-aware extraction (System.cmd in dev, Expo-based merge in releases)

Changed

  • Dashboard UI restructured from flat view to tab-based navigation
  • Improved session-based config passing between init and mount for reliable path resolution across LiveDashboard page loads
  • Updated README with full dashboard workflow documentation and screenshots

Fixed

  • Fixed gettext_path being nil on page reload by passing config through the LiveDashboard session instead of relying solely on ETS/persistent_term

[0.5.0] - 2025-11-11

Breaking Changes

LangChain Upgrade: 0.3.3 → 0.4.0

This release upgrades the LangChain dependency from 0.3.3 to 0.4.0, which introduces breaking changes in how LLM responses are handled.

What Changed

  • Message Content Structure: LangChain 0.4.0 returns message content as a list of ContentPart structs instead of plain strings
    • Before: %Message{content: "text response"}
    • After: %Message{content: [%ContentPart{type: :text, content: "text response"}]}
    • The library now automatically converts this using ContentPart.parts_to_string/1

Model Support Changes

LangChain 0.4.0 officially supports only the following LLM providers:

  • ✅ OpenAI ChatGPT - Fully supported
  • ✅ Anthropic Claude - Fully supported
  • ✅ Google Gemini - Fully supported
  • ✅ Google Vertex AI - Fully supported
  • ❌ Other models - May not work with LangChain 0.4.0

Important: If you're using ChatOllamaAI or other unsupported models, they may not function correctly with this version. We recommend:

  1. Testing your specific model after upgrading
  2. Using one of the officially supported providers
  3. Implementing a custom adapter if needed
  4. Staying on version 0.4.5 if you require unsupported models

Migration Notes

If you're upgrading from 0.4.5 to 0.5.0:

  1. No configuration changes required - Your existing configuration will continue to work
  2. Test your LLM provider - Verify translations work with your chosen provider
  3. Check the updated README - New examples added for Anthropic Claude, Google Gemini, and custom endpoints

Added

Changed

  • Updated LangChain dependency from 0.3.3 to ~> 0.4.0
  • Modified message content handling to support new ContentPart structure
  • Updated README with additional provider examples and custom endpoint configuration

Technical Details

The translate/2 function in GettextTranslator.Processor.LLM now:

  • Pattern matches on {:ok, %{last_message: %Message{content: content}}}
  • Converts ContentPart list to string using ContentPart.parts_to_string/1
  • Maintains backward compatibility with existing configurations

[0.4.5] - 2025-09-03

  • Fixed bug regarding nil values in the message strings that the Expo.PO.Composer can't handle.

[0.4.4] - 2025-06-25

  • Fixed bug with plural translations #14

[0.4.3] - 2025-06-25

  • Fixed bug regarding a new creation of changelog entries when making PRs to gitlab.

[0.4.2] - 2025-05-08

  • Fixed difference in changelog structure when making PR and saving locally to files.

[0.4.1] - 2025-05-08

  • Fixed bugs.

[0.4.0] - 2025-05-08

[Breaking chanegs]
  • Reworked changelog integration to use a simple 1:1 mapping between translations and changelog entries. Previous changelog will be overwritten, therefore new changelog entries will be created with a different structure.
  • Fixed bugs, refactored code.
  • Added visibility which changes are done in current session and will be commited or saved to files.

[0.3.0] - 2025-05-04

  • Git integration for committing changes as PRs.
  • Support Gitlab and Github (not fully tested).

[0.2.3] - 2025-05-01

  • Fixes for release environments

[0.2.2] - 2025-05-01

  • Update regex pattern in extract_language_code/1 to correctly match LC_MESSAGES folder structure
  • Extract common path patterns to module attributes to reduce repetition
  • Add private app_path/2 helper to standardize path generation
  • Ensure consistent path handling across all helper functions

[0.2.1] - 2025-05-01

Fixed

  • Fixed path resolution in release environments by using Application.app_dir
  • Added application parameter to properly resolve paths in releases
  • Fixed Ukrainian pluralization by adding support for required third plural form
  • Modified TranslationStore to correctly save and load translations in releases
  • Added ETS-based configuration for LiveDashboard integration
  • Improved PathHelper to handle both development and production environments

Added

  • New helpers for proper path resolution in both dev and release environments
  • ETS-based configuration state management between LiveView mounts
  • Support for additional plural forms required by specific languages

[0.2.0] - 2025-03-08

Added

  • Phoenix LiveDashboard integration for monitoring and managing translations
  • In-memory translation store using ETS tables
  • Web UI for viewing, editing, and approving translations
  • Filtering and pagination in dashboard
  • Improved documentation with dashboard setup guide

Fixed

  • LLM integration for gettext files.

[0.1.0] - 2025-02-23

Added

  • Initial release of GettextTranslator.
  • Multi-provider support for AI translation.
  • CLI integration for translating Gettext files.
  • Basic documentation and configuration examples.