OnecaratEditor Core — User Guide

1. Setup: preparing the model

When you first open the AI chat, Core checks whether a model is ready. If not, a setup screen guides you through one of two paths:

• Download the recommended model (Gemma 4 E4B). Because the model is license-gated on Hugging Face, you will be asked for a free Hugging Face account token.
• Use a GGUF model file you already have — just select it from disk. Your choice is remembered.

Downloads can be resumed if interrupted. Models are stored on your PC, and this setup download is the only time the app accesses the internet.

The AI runs on a local inference server (llama-server) that Core starts and stops automatically on 127.0.0.1 — traffic between the app and the model never leaves your PC. A Vulkan-capable GPU is used automatically when available.

2. Editor basics

• Tabs: open multiple files side by side. Switch with Ctrl+Tab, jump by number with Ctrl+1–8 (Ctrl+9 = last tab), middle-click a tab to close it.
• Encodings: UTF-8, Shift_JIS, and EUC-JP are detected automatically on open, and you can convert encoding and line endings (LF / CRLF) from the status bar.
• Automatic backup: your unsaved text is backed up continuously. If the app closes unexpectedly, a restore banner appears on next launch.
• View: font size 9–32 (Ctrl + mouse wheel works too), line numbers, word wrap (Alt+Z), and column (box) selection.
• Recent files (last 8) from the File menu; unsaved tabs are guarded by a save/discard/cancel dialog on exit.
• Right-click menu: cut / copy / paste / select all / AI text processing / print. The menu is Core's own — the browser-style menu (with its risky "Reload") is disabled.

3. AI chat

Open the chat sidebar with Ctrl+Shift+L or the toolbar button. Pick a character, type, and the reply streams in.

• Characters: Coashi (default assistant), Onecarat, RinRin, Biz, and You (a free-form persona you write yourself). You can also create fully custom characters — see the Characters page.
• Character portraits are shown beside the chat; some characters have alternate art you can switch by clicking.
• Turn on the RAG toggle to ground answers in your indexed folders: the reply cites its sources ([1], [2], …), and clicking a citation opens the original file in the editor.
• Chat history is kept in memory only — it is not written to disk, and clears when you close the app.

4. Folder RAG

Folder RAG lets the AI answer from your own documents. Open the RAG panel from the sidebar and add a folder:

1. Core scans the folder and shows an estimate first — how many files, how large, roughly how long indexing will take.
2. Confirm to start indexing. Progress is shown per phase (preparing the model → scanning → embedding), and you can cancel anytime.
3. When it finishes you get a summary. Files that were skipped (too large, unsupported, unreadable) are listed with reasons.
4. Re-indexing is incremental: only files that changed are processed, and an interrupted run resumes where it left off.

• Supported formats: txt / md.
• Limits: files over 5 MB are skipped; a folder is indexed up to roughly 3,000 files / 150 MB (anything beyond is listed as out of scope).
• A verification search box in the panel lets you query the index directly to see what the AI would find.

Everything — the index database, the embedding model, the search — lives and runs on your PC.

5. AI text processing

Select text in the editor, then press Ctrl+Alt+O, use the Edit menu, or right-click → "AI text processing".

• Seven tools: summarize / proofread / rephrase / translate / extract / format / generate questions.
• Each tool offers purpose and output-format presets as chips, so you shape the result in three quick picks — no prompt writing needed.
• The result streams in; you can insert it at the cursor (replacing the selection), open it in a new tab, or copy it. Press Esc to cancel a running request.

6. Voice input

Click "Voice input" on the toolbar to open the organizer card. Speak naturally — Core detects speech automatically (VAD) and transcribes each utterance locally with Whisper, adding timestamped lines as you go.

• Tabs on the card: Transcript / Summary / Meeting minutes / To-do / Ideas. The last four are generated on demand from the transcript, and each view is cached.
• Insert the result into the editor at the cursor, or open it as a new tab.
• Long recordings are supported — audio is processed in chunks, so extended sessions transcribe reliably.
• GPU acceleration is probed automatically on first use; if your GPU can't run it, Core falls back to CPU by itself.
• If the speech model is missing, the card offers to download it on the spot (one-time).

7. Split preview

Files in supported formats automatically get a live preview pane beside the editor. Toggle it with Ctrl+Shift+V or the status-bar button; drag the divider to resize.

• Markdown — GitHub-flavored tables and lists, math (KaTeX), Mermaid diagrams, syntax-highlighted code, definition lists.
• HTML — rendered in a fully sandboxed frame with scripts disabled by design; relative images and CSS resolve against the file's folder.
• CSV / TSV — parsed properly (quotes, escapes, newlines inside cells) and shown as a table with row numbers.
• JSON — pretty-printed with highlighting; SVG — sanitized and rendered.

The preview follows your edits as you type (with a short debounce).

8. Search

• Folder search (Ctrl+Shift+F): recursive text search across a folder with up to three stacked conditions; encodings are detected per file. Binary and very large files are skipped automatically.
• Big Grep (Ctrl+Shift+B): a streaming line search built for huge files — log files up to the 2 GB class — with the same three-condition filtering.
• Results are capped at 10,000 hits; click a hit to open the file at that line.

9. Built-in help

Press F1 or the "?" button and ask in plain language. Core searches the bundled help documents (getting started, editor basics, model setup, RAG usage, voice input, troubleshooting, limitations) and answers with source chips — click one to open the full help document.

If the AI model isn't running, help falls back gracefully: it will start an available model, or simply show you the matching help excerpts.

10. Keyboard shortcuts

Defaults below. All of them can be rebound in Preferences → Shortcuts by pressing the new key combination (per-row and full reset available).

Monaco's standard editing keys (find Ctrl+F, replace Ctrl+H, undo/redo) work as usual.

11. Where your data lives

• Your documents: wherever you save them — Core never uploads them anywhere.
• App data (AI models, RAG index, help index, custom character images, settings caches): %LOCALAPPDATA%\com.onecarat.editor.core.
• Chat and help conversations: memory only, never written to disk.
• Details are in the Privacy Policy.