AI Input Engine

AI Input Engine is a Windows-native input method that transparently transforms text as you type in real time, before it is committed to the target application.

Physical Keyboard → Windows Input → AI Input Engine → Target Application

Type in any Windows application (Notepad, Word, browsers, VS Code, Discord, Slack, Teams...) and AI Input Engine automatically transforms the text in-place:

1. Multilingual Live Translation: Translate across languages (English, Spanish, Chinese Simplified, Chinese Traditional, French, German, Japanese, Portuguese, Italian).
2. Same-Language Correction & Rewrite: Clean up typos, missing accents, grammar errors, and style while preserving meaning and voice in the same language.
3. Smart Paste (Ctrl+Shift+V): Instantly translate or rewrite text directly from the clipboard, synthetically paste into the active window, and seamlessly restore the user's original clipboard content.

Architecture

flowchart TD
    subgraph Trigger["1. Input & Trigger Sources"]
        direction TB
        K["Keystroke Typed<br/><i>(Global Hook / TSF)</i>"] --> IE["InputEngine<br/><i>(Buffer & Composition)</i>"]
        SP["Smart Paste Hotkey<br/><i>(Ctrl+Shift+V)</i>"] --> SPM["SmartPasteManager<br/><i>(Clipboard Snapshot)</i>"]
    end

    IE -->|"Dispatch Request"| TR["TranslationRequest"]
    SPM -->|"Dispatch Request"| TR

    subgraph Pipeline["2. Pipeline & Provider Execution"]
        direction TB
        TR --> TP["TranslationPipeline<br/><i>(Exact Cache & Monotonic ID)</i>"]
        TP -->|"Cache Miss"| PROV["OpenAI / Gemini Provider"]
        
        subgraph Worker["Worker Thread (SSE Streaming Loop)"]
            direction TB
            PROV --> SSE["• 'stream': true & SSE Parser<br/>• Reasoning effort & Delta accumulation<br/>• TTFT (Time-To-First-Token) measurement<br/>• Instant cancellation on obsolete request<br/>• Terminal completion validation"]
        end
        
        SSE -->|"Single Final Callback"| TRES["TranslationResult"]
        TRES --> TP_RESP["TranslationPipeline<br/><i>(Drop Stale / Cache Store)</i>"]
    end

    subgraph Output["3. Safe Output & Target Execution"]
        direction TB
        TP_RESP -->|"Buffer Version Guard"| IE_OUT["InputEngine<br/><b>CommitReplacement()</b><br/><i>(Atomic In-Place Replacement)</i>"]
        TP_RESP -->|"Target PID/HWND Guard"| SPM_OUT["SmartPasteManager<br/><b>Safe Synthetic Paste</b><br/><i>(Paste & Clipboard Restore)</i>"]
    end

    style Trigger fill:#1e293b,stroke:#3b82f6,stroke-width:2px,color:#f8fafc
    style Pipeline fill:#0f172a,stroke:#6366f1,stroke-width:2px,color:#f8fafc
    style Worker fill:#1e1b4b,stroke:#818cf8,stroke-width:1px,stroke-dasharray: 5 5,color:#f8fafc
    style Output fill:#1e293b,stroke:#10b981,stroke-width:2px,color:#f8fafc

Loading

Capabilities & Architecture

• Transform Operations: Authoritative Translate vs CorrectAndRewrite.
• Writing Profiles: Neutral (default), Professional, Casual, Slang, Custom.
• Supported Providers: Mock (deterministic local engine), OpenAI (SSE Streaming Responses API with structured system instructions and reasoning effort control), Google Gemini (gemini-3.7-flash, gemini-3.6-flash, gemini-3.5-flash-lite).
• Language Registry: First-class support for zh-Hans (Chinese Simplified), zh-Hant (Chinese Traditional), en, es, fr, de, ja, pt, it, and auto (for translation source).
• Security & Privacy: API keys are securely stored in Windows Credential Manager (DPAPI); never logged or stored in plaintext.
• Safety & Anti-Conversational Guarantees: Strict instruction guardrails ensure the AI only translates or rewrites and never answers questions or injects conversational filler. Preserves protected tokens (URLs, emails, mentions, hashtags, code).
• Exact Cache Identity & Monotonic IDs: Isolation key across providerId | model | operation | sourceLanguage | targetLanguage | profile | customInstructions | text. Monotonic request tracking guarantees obsolete responses are discarded.
• Speculative Translation (PCHP): Pre-Commit Hesitation Prediction preemptively streams translations during natural typing pauses, achieving near 0-ms commit latency.

Components

Component	Purpose
AIInputEngine.exe	System tray host, hook-based input engine, Smart Paste manager, modern settings & diagnostics UI
AIInputEngineTSF.dll	Text Services Framework (TSF) text service (native IME path)
aiie_unit_tests.exe	Comprehensive unit & simulation test suite (278 deterministic tests, 0 failures)
aiie_openaitest.exe	Live OpenAI connectivity, translation, streaming, and rewrite validation tool
aiie_geminitest.exe	Live Gemini diagnostic tool
aiie_tsfhost.exe	TSF host proof tool verifying composition -> transformation -> commit
aiie_tsfreg.exe	Per-user TSF registration / unregistration / verification tool
aiie_integration_test.exe	Automated Notepad hook integration test

Quick Start

Requirements: Windows 10/11 x64, Visual Studio 2026 (or 2022) Build Tools with "Desktop development with C++", CMake 3.24+.

# Build in Release mode
cmake -S . -B build -G "Visual Studio 18 2026" -A x64
cmake --build build --config Release --parallel

# Run unit tests
.\build\Release\aiie_unit_tests.exe

# Run live OpenAI smoke test matrix
.\build\Release\aiie_openaitest.exe --matrix

# Create portable distribution package
.\tools\package_dist.ps1

Hotkeys (Configurable)

Action	Default
Toggle AI Input	Ctrl+Shift+Space
Smart Paste	Ctrl+Shift+V
Swap Source ↔ Target	Ctrl+Shift+S
Diagnostics Window	Ctrl+Shift+D

License

MIT — see LICENSE.