Ghost Completion — your sentence, finished before you finish it.

You pause. A grey continuation appears. Tab — accepted. Esc — gone. Knovya's inline AI autocomplete is native to the editor, knows your knowledge base and your block type, and stays quiet inside encrypted blocks. Off by default; on Pro across every surface — paragraph, heading, list, table cell, fenced code block, voice transcript. See the full editor on AI Co-Edit or compare plans on pricing.

Keys (Tab · Esc)
2
Block types it knows
4
Algorithmic ancestors
5
Ghost Completion
Experiment 01 · The Lab

Type a line. Watch Ghost finish it.

Press play. Ghost shows up shortly after the cursor pauses, in a dimmed grey italic. Tab to accept, Esc to dismiss. Three presets, one editor, every keystroke deterministic.

Auth migration plan — draft
Ghost on
The pipeline · what happens between keystrokes

Four stages, silently.

Every suggestion travels four steps before the grey text reaches your eye. Knowing the pipeline is knowing why Ghost stays out of the way when it should.

  1. 1

    Sense

    When to fire — and when not to.
    Pause detection

    Ghost waits for a brief lull in your typing — not the comma, the breath.

    Dismiss-on-type

    The moment you keep typing, the suggestion is gone. No fight for the cursor.

    Cooldown after a reject

    Esc is heard. Ghost backs off briefly before trying again.

  2. 2

    Read

    Just enough context — never more.
    Recent text before the cursor

    Your last few sentences shape the next one. Older paragraphs do not pollute the prompt.

    Block type, title, tags

    Ghost knows whether you are in a heading, a list item, a table cell, or a code fence.

    Encrypted notes are silent

    E2E-encrypted blocks never enter the prompt. Ghost is mute inside them, by design.

  3. 3

    Predict

    A fast model, on a tight leash.
    Fast tier, small response

    Speed beats verbosity. The model returns one or two sentences — never a paragraph you did not ask for.

    Fair-use limits

    Burst protection prevents runaway calls when you are writing in flow. The pacing is invisible.

    Cancel on keystroke

    If you start typing while a request is in flight, the response is dropped before it lands.

  4. 4

    Show

    Dimmed, dismissable, deliberate.
    Single or multi-line

    If the prediction crosses a line break, Ghost goes multi-line. Otherwise it stays inline.

    Tab accepts, Esc dismisses

    Two keys, no menu. Accept the whole suggestion or take a single word with one shortcut.

    Three-strike guard

    Three rejections in a row pauses Ghost for the rest of the session. The signal is loud and clear.

Most AI helps after you finish writing —
but the hardest moment is the next word.

Open a chat panel, paste a paragraph, prompt for a rewrite, paste it back. Tab-out, tab-in, attention spent.

The hardest moment in writing is the next word — and that is the one moment most AI tools are not there for. The prompt window is too far away. The cursor is not.

The cost
Every prompt-and-paste cycle pulls you out of your own sentence. Cognitive switching tax, paid in flow.
The fix
Show the next word in place. Dim it so it is dismissable. Trust two keys.
The lineage

From IntelliSense to your second brain.

Ghost Completion is not invented from nothing. Five tools across thirty years taught it how to step in — and how to step aside.

  1. 1996
    Microsoft IntelliSense Visual Studio's autocomplete drop-down — statistical, dictionary-based. The idea that the IDE could finish a token took root here. IDE · syntax-aware
  2. 2018
    Gmail Smart Compose Tab-to-accept, in grey, on consumer prose. The first time most writers met a prediction that knew them. Google · email at scale
  3. 2021
    GitHub Copilot LLM-powered ghost text in the code editor. Multi-line, fill-in-the-middle, dismissable. Showed that prediction could be useful, not just clever. GitHub × OpenAI · IDE
  4. 2023
    Cursor Anysphere forked VS Code to make ghost completion repository-aware. Proved the next breakthrough was in context, not just the model. Anysphere · AI-native IDE
  5. 2026
    Knovya Ghost Completion Native to a knowledge base. Block-type aware, tag-aware, encrypted-block-silent. The first ghost that knows what kind of writing it is interrupting. ★ Knovya · production
First of its kind

Nobody else writes the next sentence in your notes.

Notion AI waits for you to highlight and ask. Obsidian needs a community plugin. Mem and Reflect put AI in a chat panel. Copilot and Cursor are for code. None of them know your knowledge base, and none of them know what kind of block you are in.

  • Notion AI request-based · highlight + ask
  • Obsidian community plugin required
  • Mem chat panel · no inline ghost
  • Reflect chat panel · no inline ghost
  • GitHub Copilot code only
  • Cursor code only
  • Smart Compose email only
  • ★ Knovya native · KB-aware · block-aware
Surfaces

Ghost adapts to whatever you are writing.

Block-type awareness is not a setting. It is what Ghost does. The same model behaves differently inside a paragraph, a list, a code fence, or a voice transcript.

Paragraph prose

default

The home surface. Continues your sentence with the rhythm and tone of the surrounding text — never a paragraph longer than you asked for.

The migration plan moves auth from Cognito to a self-hosted OAuth flow, while keeping fallback to Cognito for the first sixty days. Block: paragraph · Suggestion: 12 words · Tab to accept

Heading + list

block-aware

In a heading, suggestions are short and noun-led. In a bulleted list, Ghost continues the pattern instead of breaking it.

Open questions for next week Decide on token rotation cadence Lock the rollback plan Schedule the security review with the platform team

Code block

language-aware

Inside a fenced code block, Ghost switches to language-aware suggestions — JavaScript, Python, Rust, SQL. Not a chat panel, not a separate IDE.

typescript
function refreshToken(token: string)
  if (isExpired(token))
    return await rotate(token, { fallback: true });
  

Voice transcripts

cleanup mode

When a voice note becomes prose, Ghost helps polish — fixing the half-finished sentence the speaker abandoned, completing the thought they restarted.

recorded · 1m 24s
"…and the customer basically said the onboarding was, um, too long and they—" they didn't make it past the second screen.
Bonded with

Ghost composes with the rest of Group I.

03
AI Co-Edit
The drawer chat layer. Ghost is its passive twin — both inside one editor.
04
AI Transforms
The selection layer. Highlight, transform — Ghost predicts, Transforms revises.
02
AI Memory
The shared brain Ghost reads when the surrounding paragraph is not enough.
07
Voice Notes
A second surface for Ghost — finishing the sentence the speaker abandoned.
Frequently asked

A few honest answers.

What is Ghost Completion in Knovya?
Ghost Completion is Knovya's inline AI writing assistant. As you type, it predicts what you are likely to write next and shows a dimmed grey continuation at the cursor. Tab accepts it, Esc dismisses it. It is block-type aware (knows whether you are in a heading, list, table, or paragraph) and knowledge-base aware (uses the rest of your notes as context).
How is Knovya's Ghost Completion different from GitHub Copilot?
GitHub Copilot is built for code editors — it predicts the next line of code, trained on public repositories. Knovya's Ghost is built for notes — it predicts the next sentence in your prose, your bullet, your table cell. It uses your knowledge base as context, not the public web.
Is Ghost Completion on by default?
No. Ghost is off by default. You opt in per note or per workspace from Settings → Editor → Ghost Completion. The deliberate choice protects writers who want a quiet page, and it is reversible at any time with a single toggle.
What happens if I keep dismissing the suggestions?
Three rejections in a row pauses Ghost for the rest of your session. The signal is clear — the model is not in tune with what you want right now. The next session restores it. No retraining, no nagging.
Does Ghost Completion read my whole knowledge base?
Ghost reads the surrounding text, the block type, the note title, and a small set of tags. It can also draw on linked notes when the context warrants. End-to-end encrypted notes are never read — Ghost is silent inside encrypted blocks, by design. See Privacy & Security.
Does Ghost Completion work in code blocks?
Yes. Ghost is block-type aware. In a fenced code block, it switches to language-aware suggestions. In a heading, it suggests short titles. In a bulleted list, it continues the pattern. In a table cell, it stays terse.
Is Ghost Completion free?
Free includes a daily quota of basic prose suggestions. Pro unlocks full block-aware and knowledge-base-aware Ghost across every surface — paragraph, heading, list, code, voice transcript. See Pro.

Try Ghost in your own notes.

Off by default. On with a click. Two keys, one editor. Your sentence, finished only when you say so.

Start free See the full editor
★ Element 08 · Group I — AI · /features/ghost-completion