Jingwen Wang70 downloadsFact-check a selected passage with an LLM (DeepSeek, Gemini, Claude...) and write corrections back in place with a particle animation. Keeps your own wording.
Select a passage in a note and let an LLM check it for factual errors. Your own wording is kept; only what is actually wrong gets fixed. The corrected passage dissolves into ink particles and re-forms in place, changed spots are marked with a soft gray ==highlight==, and a short teacher-style note appears right under the passage: what you got right, or what was wrong and why.
中文说明见下方。
== markers in the selection and keeps the text.Ctrl/Cmd+Z undoes a whole review in one step. With no selection, nothing happens.While waiting, the selection shimmers softly. When the result arrives:
Esc to skip the animation.The command palette also has Review (facts only), Review with extra instructions, Remove highlights in selection, Cancel the running review, and Demo animation (no API call) (inserts a sample passage and plays the correction without calling any API; one Ctrl+Z removes it).
jiwawang/obsidian-note-proofreader. BRAT keeps it updated.main.js, manifest.json and styles.css from Releases into <vault>/.obsidian/plugins/note-proofreader/, then enable the plugin.Requires Obsidian 1.13.0 or later. After enabling, pick a provider and add your API key in the plugin settings.
The default is any OpenAI-compatible endpoint, with presets for inexpensive options:
| Preset | Default model | Rough price (per 1M tokens, in / out) | Notes |
|---|---|---|---|
| DeepSeek | deepseek-flash |
$0.15–0.30 / $0.60–1.20 | Default. Good in Chinese and English; a review costs a fraction of a cent. |
| Google Gemini | gemini-2.5-flash-lite |
$0.10 / $0.40; Flash models have a free tier | On the free tier your text may be used for training. |
| OpenRouter | openrouter/free |
free (random free models) | Quality varies; any model ID works. |
| Custom | — | — | Any OpenAI-compatible endpoint. |
You can also switch to Anthropic Claude (default Claude Opus 5; more expensive, more reliable checking).
| Setting | What it does |
|---|---|
| Provider / API key | OpenAI-compatible endpoint or Anthropic Claude. |
| Preset / base URL / model ID | For OpenAI-compatible endpoints. |
| Effort / fallback model / custom URL | Claude only. |
| Context length | How many characters of the surrounding note are sent for context. |
| Double-tap Alt / interval | Turn the shortcut off, or tune how fast the two taps must be. |
| Feedback card duration | How long the card stays; 0 keeps it until you close it. |
| Highlight colour | How every ==highlight== in your vault looks: gray (default), light blue, or the theme's own colour. |
| Particle animation | Master switch. Animation timings are fixed in src/params.ts. |
npm install
npm run build
npm run dev rebuilds on change.
MIT
选中一段笔记,让大模型检查其中的事实性错误。以你的原话为准,只改确实错的地方。修正后的段落会化成墨色颗粒、再在原处重新聚成;改动处写入浅灰色 ==高亮==(设置里可改为浅蓝或主题默认色);段落正下方出现一张毛玻璃卡片,像老师一样告诉你:哪里对、对在哪,或者哪里错、为什么错。
== 标记,保留文字。Ctrl/Cmd+Z 一步还原。没有选中文字时什么都不做。