Skip to content

SubBuild Studio v2.0.1

Choose a tag to compare

@talevi83 talevi83 released this 18 Nov 22:52
· 35 commits to main since this release

SubBuild Studio v2.0.1 — Multi-Provider AI Workbench, GPU-Aware Cost Optimizer & Built-In Auto-Updater

SubBuild Studio v2.0.0 unifies the v2.0 series into a complete Multi-Provider AI Subtitle & Translation Workbench—combining multi-provider LLM and vision routing across Google Gemini, OpenAI, Anthropic Claude, and Local / Cloud Ollama, hardware-aware cost/value model optimization, a 2-column Bento settings layout, a source-first desktop launcher, and a built-in GitHub release auto-updater.

What's New

1. Multi-Provider AI Architecture & Unified LLM Routing

  • 4 AI Providers Across Every Pipeline Stage: Mix and match Google Gemini, OpenAI, Anthropic Claude, and Ollama (Local GPU or Ollama Cloud) independently across:
    • Subtitle Translation
    • Series Glossary & Character Gender Discovery
    • AI Vision / On-Screen Text (Forced Narratives OCR), Speaker Detection & Sync Line Matching
  • Multi-Account API Key Vault (storage/provider_keys.json): Store, label, test, enable/disable, and set default API keys per provider directly in the UI, with custom base URL support, one-click Ollama Cloud preset configuration, and secret masking across API responses and view reloads.
  • Live Model Discovery & Capability Filtering: Queries live provider endpoints on startup and key refresh (GET /v1/models, Anthropic models API, Gemini models API, and Ollama /api/tags + /api/show), deduplicates dated OpenAI snapshots, verifies local Ollama vision capabilities via /api/show, and filters stage dropdowns to verified, key-backed, task-capable models.

2. One-Click "Load Best Cost / Value Models" (GPU & VRAM Aware)

  • Smart Auto-Recommender: Evaluates configured API keys to recommend cost-effective, high-quality models per stage, factoring in 50% Google Flex discounts when comparing cloud pricing.
  • Hardware-Gated Local AI (Include Local AI): Optionally includes $0 token-cost local Ollama models based on detected GPU/VRAM (e.g., 8 GB vs. 12 GB+ vs. CPU), skipping oversized models (such as 14B / 20B+ on 8 GB GPUs) with clear suitability badges.

3. Deduplicated Bento Settings & Studio / Ingest Enhancements

  • 2-Column Bento Settings Grid:
    • Row 1: Appearance (Language & Theme) alongside a compact Local AI & Cost-Saver status card with a collapsible setup drawer.
    • Row 2: Streamlined API Key Vault with provider tabs, inline connection status, and quick-apply actions.
    • Row 3: Side-by-side 3-Column AI Stage Cards (Translation, Glossary, Vision) with 45m episode and 120m feature-film cost badges, suitability warnings, and a collapsible Interactive Cost Estimator.
    • Row 4: Storage Cleanup alongside Diagnostics & Feedback (server log download, bug report, and feature request shortcuts).
  • Cleaner Job Presets: Moved job-specific toggles (Metric Conversion and Remove SDH Speaker Labels) out of global Settings into the post-video Pipeline Configuration screen.
  • Studio, Ingest & Localization: Sidebar and ingest summaries of active providers across stages with missing-key warnings, capability-filtered row re-translation with provider-aware Thinking controls, and full EN / HE (RTL/LTR) localization.

4. Built-In GitHub Release Auto-Updater (VERSION, backend/app/core/updater.py)

  • Pre-Launch Release Check (run.bat / run.sh): Checks origin for newer semantic release tags (vX.Y.Z) before starting the server via python -m app.core.updater --auto --no-frontend-build, ignoring unreleased branch commits.
  • Developer & Offline Safe: Network failures never block startup; --auto skips updates on feature branches (feature_branch) or when tracked files have uncommitted local edits (dirty_worktree), using --ff-only merges.
  • Automatic Sync & Opt-Out: Updates Python dependencies when backend/requirements.txt changes, rebuilds frontend/dist when frontend sources change, and supports --no-update or SUBBUILD_NO_UPDATE=1.
  • In-App Version Badge & Diagnostics UI: Displays the installed vX.Y.Z version and update notification badge in the sidebar footer, plus Settings -> Diagnostics -> Software Version & Updates (POST /api/system/update/check and POST /api/system/update/apply) with a 409 Conflict guard while pipeline jobs are active and audit logs in storage/logs/server.log.

5. Source-First Desktop Window & Removal of PyInstaller .exe Packaging

  • Retired Legacy Bundle Files: Removed build_exe.bat, packaging/subbuild.spec, packaging/requirements-build.txt, packaging/get_ffmpeg.ps1, and packaging/ffmpeg.sha256.
  • Native Desktop Launcher (packaging/launcher.py): Runs directly from source via run.bat (Windows) or ./run.sh --window (macOS/Linux) using Edge WebView2 (pywebview), pystray system tray integration, zero-copy local file path drag-and-drop, and automatic packaging/ffmpeg PATH resolution—avoiding 2–3 GB binary rebuilds.

6. Docs, Configuration & Automated QA Suite

  • Updated Documentation: Refreshed README.md, SPEC.md, and backend/.env.example to clarify that multi-provider API keys and per-stage models live in Settings -> API Key Vault (storage/provider_keys.json & storage/settings.json), while backend/.env remains optional for low-level server and Whisper hardware defaults, alongside updated Bento settings, studio tools, and feature/* -> main PR workflow docs.
  • Test Coverage: Verified with 512 Pytest backend tests (including updater unit tests and active-job 409 tests), 14 Vitest frontend tests, and 53 Java 21 (Selenium 4 + RestAssured) E2E/API tests.