Lexicon v0.6.0

v0.6.0 - Major Landmark Release: 4-Step Local Onboarding & Complete App Overview

Quick Downloads (Most Popular Releases):

• 🪟 Windows x64 Setup (Most Popular): Standard installer for modern 64-bit Windows PCs (Intel / AMD).
• 🍏 macOS Apple Silicon DMG (Most Popular): For modern Apple Silicon Macs (M1, M2, M3, M4 chips).
• 📦 View All 15 Platform Assets & Checksums: Complete list of installers including ARM64 Windows, x86 Windows, and Intel macOS.

Note

🛡️ First-Launch Security & Verification Notice

Why am I seeing a security warning?
Lexicon is 100% free, private, and open-source software built directly from source code. Code Signing certificates required by OS vendors are expensive ($$$+/yr), Windows Defender and macOS Gatekeeper will show a standard "unrecognized app" warning on first launch. Lexicon is completely safe, offline, and open-source.

• 🪟 Windows (SmartScreen):
  If Windows shows "Windows protected your PC", click "More info" $\rightarrow$ click "Run anyway".

• 🍏 macOS (Gatekeeper):
  If macOS shows "Lexicon cannot be opened because it is from an unidentified developer":

  • Option 1 (GUI): Control-click (or Right-click) Lexicon.app $\rightarrow$ click Open $\rightarrow$ click Open.
  • Option 2 (System Settings): Open System Settings $\rightarrow$ Privacy & Security $\rightarrow$ scroll down and click "Open Anyway".
  • Option 3 (Terminal): Open Terminal and run:

    xattr -cr /Applications/Lexicon.app

⚠️ Troubleshooting
If you experience a blank white screen on startup or any other issues:

• Windows: Click the System Tray chevron (^) near your clock, right-click the Lexicon icon, select Quit Lexicon, and relaunch after a few seconds, give it a few seconds to fully close!
• macOS: Right-click (or Control-click) the Lexicon icon in your Menu Bar (top-right) or Dock, select Quit, and relaunch after a few seconds.

---

Welcome to Lexicon v0.6.0!

Lexicon is a private, local-first writing assistant designed for calm, distraction-free drafting. Everything runs directly on your device; no accounts, no cloud subscriptions, no tracking, and zero data leaving your machine.

Key Features Included:

• 100% Offline & Private: Your drafts, notes, and documents stay strictly on your local hardware.
• Local Grammar & Spellchecking: Instant, deterministic proofreading powered by local LanguageTool (zero LLM latency).
• Your Local Assistant (Lex) 'J': Opt-in local AI for rewriting, tone adjustments (Friendly, Professional, Academic, Formal, Casual, Playful, Empathetic, Persuasive, Humorous), and document summaries running entirely on your machine or an optional Ollama server.
• Distraction-Free Workspace: Rich-text editing with headings, lists, blockquotes, typography rules, slash commands, LaTeX math ($E=mc^2$), resizable side panels, and Focus Mode.
• Import & Export: Support for .md, .txt, .html imports and clean PDF / Markdown exports.

---

🐛 Feedback & Support

Encountered a bug, crash, or unexpected behavior? I'd love to fix it!
Please report any issues or submit feedback on the GitHub Issues Tracker.

---

What's New in v0.6.0:

• 4-Step Local-First Onboarding Experience: Interactive wizard (OnboardingModal.jsx) guiding new users through Lexicon's privacy philosophy, language dialect preferences (en-US, en-GB, etc.), Lex AI assistant configuration, and sample draft ingestion.
• Interactive Sample Document: Pre-populated onboarding draft featuring live proofreading squiggles, LaTeX math ($E=mc^2$), and AI rewrite prompts.
• UI Icon & Aesthetics Refinement: Phosphor SVG icons (ShieldCheck, PencilLine, Robot, Confetti), branded logo visuals, and clean slate paper aesthetic.
• Process & Type Safety Patches: Subprocess creation window suppression with __class_getitem__ type hint safety for Windows sidecar execution.
• Shared System Prompt Parity Across Inference Backends: Unified SYSTEM_PROMPT constant across bundled GGUF models and external Ollama servers (v0.5.19).

Lexicon v0.5.19

v0.5.19 - Shared System Prompt Parity Across Inference Backends

Changes Made:

• Unified System Instructions: Defined a single module-level SYSTEM_PROMPT constant in backend/inference.py enforced across both local bundled GGUF models and external Ollama servers.
• Ollama System Field Inclusion: Passed top-level "system": SYSTEM_PROMPT to Ollama /api/generate requests, preventing external models (llama3, mistral, deepseek-r1, qwen2.5) from outputting conversational preamble or reasoning chatter.
• Architectural Prompt Parity: Refactored BundledBackend and OllamaBackend to reference the single source of truth constant, guaranteeing complete behavioral parity and preventing prompt drift.

Lexicon v0.5.18

v0.5.18 - Non-Blocking Startup, Instant Quit & Release Workflow Prioritization

• Non-Blocking UI Startup: Backend sidecar startup runs on an asynchronous background thread so the window and index.html load instantly (< 1ms) without blocking the OS UI thread.
• Instant Window Hiding on Quit: Main window hides (window.hide()) on line 1 of the Quit handler so the desktop UI vanishes instantly (< 1ms) while process termination completes silently in the background.
• Native Window Theme & Boot Optimization: Configured "theme": "Light" and "backgroundColor": "#f7f7f5" in tauri.conf.json (matching index.html), and set Google Fonts to non-blocking loading for instant frame-1 boot screen rendering.
• Detached Process Cleanup & Console Suppression: Added silent Java process termination on quit (taskkill /F /IM java.exe) and global Windows CREATE_NO_WINDOW / SW_HIDE subprocess overrides to eliminate console window flashes.
• Release Workflow Prioritization: Updated .github/workflows/release.yml to prioritize extracting release notes directly from CHANGELOG.md first.

Lexicon v0.5.17

v0.5.17 - Model Memory Mapping, Context Bounds & Paragraph-Level LRU Caching

• Model Context Window Bounds (n_ctx = 4096): Reduced local LLM context window from 8,192 to 4,096 tokens, cutting KV-cache RAM/VRAM allocation by 50% (~300–600 MB savings) and doubling prompt evaluation speed.
• Memory Mapping (use_mmap = True): Configured OS-level memory mapping for GGUF model weights for near-instant model loading (< 0.5s) and instant physical RAM page reclamation during Tier 1 offloads.
• Frontend Chunking Budget Alignment: Updated TRANSFORM_INPUT_BUDGET and TRANSFORM_CHUNK_BUDGET to 1,800 tokens (~7,200 chars / ~3–4 full paragraphs), with 2,048 max output token headroom.
• Sub-Microsecond 64-Bit Hashing: Implemented synchronous fnv1a64(str) hashing (< 20 nanoseconds per block) for instant LRU cache key calculations without Promise overhead.
• Context-Aware Predecessor Keys: Paragraph cache keys incorporate the preceding paragraph's suffix (suffix(A, 64)). Deletions, additions, or merges automatically update predecessor keys and invalidate downstream boundary rules without requiring complex cascade-invalidation code.
• Differentiated Proofread Triggers: Live typing auto-checks evaluate only modified/uncached paragraphs; clicking the manual Proofread button or pressing Ctrl/Cmd + Enter bypasses the cache to perform a 100% full-document scan.
• Non-Blocking UI Startup: Backend sidecar startup runs on a background thread so the window and index.html load instantly (< 1ms) without blocking the OS UI thread.
• Instant Window Hiding on Quit: Main window hides (window.hide()) on line 1 of the Quit handler so the desktop window vanishes instantly while process cleanup completes in the background.
• Native Window Theme & Boot Optimization: Configured "theme": "Light" and "backgroundColor": "#f7f7f5" in tauri.conf.json, and set Google Fonts to non-blocking loading in index.html for instant frame-1 boot screen rendering.
• Detached Process Cleanup & Console Window Suppression: Added silent Java process termination on quit (taskkill /F /IM java.exe) and global Windows CREATE_NO_WINDOW / SW_HIDE subprocess overrides to eliminate console window flashes.
• Release Workflow Prioritization: Updated .github/workflows/release.yml to prioritize extracting release notes directly from CHANGELOG.md first.

Lexicon v0.5.16

Lexicon v0.5.16

Changes/Updates:

• Windows Job Object Cleanup: Bound backend sidecar processes to a OS-level Job Object (JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE), ensuring lexicon-backend.exe and java.exe are automatically killed when Lexicon is terminated via Task Manager End Task.

• Silent Background Pre-Warming: Clicking into the editor canvas or focusing the window silently wakes offloaded backend tiers in the background before typing completes.

• Aggressive Heap Bounds: Capped JVM heap memory to -Xms64M -Xmx384M.

• G1GC Memory Return Tuning: Configured -XX:+UseG1GC with -XX:MinHeapFreeRatio=10 and -XX:MaxHeapFreeRatio=20 to force Java to immediately release unused memory back to the host operating system.

• String Deduplication: Enabled -XX:+UseStringDeduplication to reduce RAM footprint from repetitive dictionary strings and rule patterns.

• Locale Filtering: Initialized LanguageTool specifically for the user's preferred locale (en-US, en-GB, etc.), skipping unneeded language modules.

• Markdown Release Notes: Update modal renders release notes using marked with styled HTML formatting for headers, lists, bold text, and code blocks.

• Automated Updater Manifests: Updated the GitHub Actions release workflow to extract release notes into latest.json.

Lexicon v0.5.15

Lexicon v0.5.15

Changes:

1. Windows Job Object Integration (main.rs):

   • Added low-level Windows kernel Job Object bindings (CreateJobObjectW, SetInformationJobObject, AssignProcessToJobObject) with JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE.
   • Every spawned instance of lexicon-backend.exe is bound to this Job Object upon startup.
   • Result: If Lexicon.exe is terminated via Task Manager "End Task", crashes, or exits unexpectedly, the Windows kernel OS immediately and atomically terminates lexicon-backend.exe and any child java.exe processes, eliminating orphaned background processes.

2. Silent Focus Pre-Warming (api.js & App.jsx):

   • Exported ensureBackend() from api.js.
   • Added silent background pre-warming event handlers to both TipTap editor canvas focus (handleFocus) and window focus (window.onFocus).
   • Result: When returning to Lexicon after an idle period, offloaded tiers begin re-warming in the background before the user finishes typing their first sentence, completely silently with zero UI toasts or notification banners.

Lexicon v0.5.14

Lexicon v0.5.14

Privacy-first grammar and AI writing assistant. Runs locally on your computer; the local model downloads on first use.

Lexicon v0.5.14

• Tier 1 (5 Minutes Idle): Unloads LLM model weights from RAM (~1.4 GB freed).
• Tier 2 (15 Minutes Idle): Terminates the LanguageTool JVM (java.exe) (~400–700 MB freed).
• Tier 3 (30 Minutes Idle): Shuts down the sidecar process (lexicon-backend.exe).
• Resuming writing at any time transparently re-engages the required tier on demand.

Downloads

• Windows x64: 64-bit Intel/AMD Windows
• Windows ARM64: Windows on ARM; proofreading is supported, but the optional native AI engine is not bundled
• Windows x86: legacy 32-bit Windows; proofreading is supported, but the optional native AI engine is not bundled
• macOS Intel: Intel Macs
• macOS Apple Silicon: M-series Macs

macOS builds are unsigned. On first open, right-click the app and choose Open, or run:
xattr -dr com.apple.quarantine /Applications/Lexicon.app

On first launch the app downloads the local AI model (~1.4 GB for the 2B tier). Proofreading works immediately without it.

Lexicon checks for updates in the app. Downloaded models and app data are kept outside the installation directory.

Lexicon v0.5.13

Lexicon v0.5.13

Privacy-first grammar and AI writing assistant. Runs locally on your computer; the local model downloads on first use.

• Black Screen Recovery
• Terminal Window Flash fix
• Process Tree Cleanup

Lexicon v0.5.12

Lexicon v0.5.12

Privacy-first grammar and AI writing assistant. Runs locally on your computer; the local model downloads on first use.

Lexicon v0.5.12

• Added release notes modal dialog for in-app updates
• Added live progress bar indicator during update downloads
• Accelerated macOS Intel build times (~16 min -> ~5 min)
• Improved Windows file lock handling during background app updates

Downloads

• Windows x64: 64-bit Intel/AMD Windows
• Windows ARM64: Windows on ARM; proofreading is supported, but the optional native AI engine is not bundled
• Windows x86: legacy 32-bit Windows; proofreading is supported, but the optional native AI engine is not bundled
• macOS Intel: Intel Macs
• macOS Apple Silicon: M-series Macs

macOS builds are unsigned. On first open, right-click the app and choose Open, or run:
xattr -dr com.apple.quarantine /Applications/Lexicon.app

On first launch the app downloads the local AI model (~1.4 GB for the 2B tier). Proofreading works immediately without it.

Lexicon checks for updates in the app. Downloaded models and app data are kept outside the installation directory.

Lexicon v0.5.11

Lexicon v0.5.11

Privacy-first grammar and AI writing assistant. Runs locally on
your computer; the local model downloads on first use.

Downloads

• Windows x64: 64-bit Intel/AMD Windows
• Windows ARM64: Windows on ARM; proofreading is supported,
  but the optional native AI engine is not bundled
• Windows x86: legacy 32-bit Windows; proofreading is supported,
  but the optional native AI engine is not bundled
• macOS Intel: Intel Macs
• macOS Apple Silicon: M-series Macs

macOS builds are unsigned. On first open, right-click the app and
choose Open, or run:
xattr -dr com.apple.quarantine /Applications/Lexicon.app

On first launch the app downloads the local AI model (~1.4 GB for
the 2B tier). Proofreading works immediately without it.

Lexicon checks for updates in the app. Downloaded models and
app data are kept outside the installation directory.