Back to blog
Create
Product
Ask
Changelog
Version history, remote MCP and speaker notes
Features
- Version history gained diffs and names: compare any two versions, name them, and open a version read-only full-screen in the workspace; note cards in history carry a vN badge that opens that exact version.
- Custom deck templates are back in the new generation flow: upload a pptx to extract a template, preview its pages, delete it, and watch the extraction stages.
- Decks support speaker notes: gathered at generation, editable, and included in playback and PPTX export.
- Generated decks can be revised page by page — only the named page changes and every other page is preserved; uploaded .pptx files can be edited in the conversation too; a full restructure continues on the same deck instead of starting over.
- Style and page count sit in one stepped card, skipping what you already chose; the style tab shows 18 previews again, with a full gallery.
- External agents can connect over remote MCP with OAuth — confirm once on the consent page instead of copying tokens; the skill can save selected message ranges as one transcript.
Improvements
- Uploads and files are smoother: duplicates are flagged before upload, choosing overwrite updates the file in place and keeps the old version in history; uploads go direct to storage; opened PDFs and images are cached, so a refresh no longer downloads them again.
- Notes support LaTeX (including AI-written ones), and list previews stop showing raw $ signs; asking the AI to revise a note updates it in place instead of creating a second copy.
Fixes
- Chinese PDFs no longer lose or garble their text layer; pages with missing font data go to OCR instead of being stored as garbage.
- Captcha interstitials are no longer saved as successfully parsed materials; Baidu Netdisk links now say they are unsupported instead of pretending to succeed.
- A layout-only template no longer silently produces an empty deck; asking again mid-generation no longer quietly merges into the previous job; the current material refreshes after a page edit; a deleted material shows an error instead of a permanent skeleton.
- Answers written only in the reasoning channel no longer show as empty, and failed artifacts are no longer promised as successful.
Silent-video transcription, diagrams and live generation progress
Features
- Videos without captions can be transcribed on this machine: you choose the language, progress shows as it goes, and the result is clickable and selectable like a caption; captions can also be selected and sent to the conversation.
- Pasting a link into the chat imports it as a material without duplicates; public Google Drive links import directly, and private ones say they are not accessible; the import box also accepts forms like www.x.com.
- A new diagram material: diagrams fill the workspace, pan and zoom, and show live progress while generating.
- Deck generation moved to a new pipeline: a live timeline, action words and characters per second while you wait, and a draft page in the workspace; extra requests queue instead of all running at once.
- Before generating a deck the assistant asks for style and page count, skipping anything you already stated; one message can request several decks, mind maps or diagrams; cancelling really stops generation.
- Presenting and exporting decks no longer require a membership; importing, creating materials and collecting agent sessions now do, with a prompt when it expires. Generating a deck needs at least $0.14 of credit, with a top-up link in the chat when it is short.
Improvements
- Leave a material or the page and come back: parsing progress resumes; the "all materials" row stays pinned while scrolling.
- A refresh no longer re-downloads a PDF; originals and previews reuse the cache.
- Running out of credit now distinguishes an expired membership from a used-up balance and links to your subscription; the member page shows token debt and a way to buy more, with amounts in local currency.
- English users no longer see Chinese error messages (29 codes added), and English questions no longer drift into Chinese answers; Chinese copy uses the polite 您 throughout.
Fixes
- A conversation title went blank when the model returned nothing — it now falls back to your own words; failure messages no longer expose "bridge not ready" or absolute log paths.
Workspace redesign, audio/video materials and Word export
Features
- The desktop workspace dropped its top bar: the logo, search and account now sit in the left icon rail, and collapsing the right pane leaves a 44px strip you can reopen at any time. The new buttons became one, with new notes under the ▾ menu and a new-note tab in the import dialog. ⌘F / Ctrl+F now opens a dedicated search palette: search history and materials on the left, AI conversation hits in their own column, with matches highlighted and counted.
- Materials keep version history: the "···" menu lists snapshots from before an overwrite, lets you preview one read-only and restore it; continuous saves merge, up to 10 versions (3 for very large bodies).
- Video materials show as a player plus subtitles: click a caption on YouTube or Bilibili to jump to that moment; importing a link fetches the captions instead of a bare page shell; videos without captions say so. Audio materials have their own view — player plus transcript — and long audio no longer loses its opening or fixes the recognition language to Chinese.
- Direct links import by type: PDFs keep the original file and open as PDFs, while Office files, images, audio and plain text follow the upload path; HTML files import as rich-text notes. Importing shows the current stage and elapsed time, and rate limits are stated plainly.
- Text materials export to editable Word (free, like PDF); you can also ask the AI to export and get an export card in the conversation.
Improvements
- When you are on another page, the browser tab icon shows AI status: a badge while running, one for done or failed, cleared when you come back to the tab.
- Reopening a very long conversation paints in 1.3s instead of 21.7s; asking about a long document page by page no longer rereads the whole file each time (8.1s down to 0.6s).
- Multi-word search now requires every word to match — the old rule treated the whole query as one phrase and almost never found anything. English error copy, empty previews and relative times in the library are translated too.
Fixes
- Uploads and parsing: filenames with accents, kana, Korean, Arabic or Cyrillic no longer fail; the UI no longer sticks on "parsing"; image-only decks open instead of failing; scanned text stays in reading order, including two-column pages.
- Formulas: PDFs no longer break each display line into its own block or emit pages bottom-up; notes render formulas again in rich text, Markdown preview and web imports, including the \( \) and \[ \] forms.
- The AI wrote its thinking into answers and called itself just a Q&A assistant; decks overstated page counts and invented business context. Those claims and receipts are tightened.
PDF citation highlights, sign-in and material export
Features
- PDF materials switch between "original PDF" and "parsed text": click a citation and it jumps to the right place in the PDF and highlights the passage.
- The workspace "···" menu was tidied up: a leaner view menu, plus a new export-by-type entry (sharing and exporting covers every material type).
- The library gained system folders for LLM conversations and agent sessions, with a "usage guide" shortcut.
- Google sign-in works end to end and WeChat QR sign-in is back; if an email already exists, you can sign in with the password you just typed.
- A bilingual Terms of Service page, plus machine-readable /llms.txt and /pricing.md.
- A product overview plus four new articles; article bodies now render tables, figures and code blocks, with a copy button.
Improvements
- Organize plans show progress (create folders, move files) and keep their state across refresh and history.
- The English site shows the local amount PayPal actually charges; upload progress shows a real percentage instead of stalling at 10%.
- Material titles in import notices are now links that open the material in the workspace.
Fixes
- Citation numbers out of range, silent tool failures, and original files stuck on a spinner.
A new deck engine, graphs, and the blog
Features
- Deck generation moved to a new pipeline, ppt-master: pages render in parallel, and an 8-page deck went from 649s to about 200s with the same quality gate.
- Decks can be edited by hand, or revised one page at a time with one sentence — every other page is preserved word for word.
- Knowledge graph: cross-material linking can be re-run per material with visible progress, backfills a folder or everything at once, and clicking a node highlights its immediate neighbourhood.
- A new "LLM chat" tab in the sidebar and a calendar view; conversation search now covers message bodies and highlights keywords.
- The blog shipped: article list, article pages and this changelog.
Improvements
- Graph extraction got more than twice as fast (a 20k-character material went from 31s to 9s).
- Collapsing the left pane keeps the icon rail; a refresh restores where you were (conversation, calendar month, graph range).
- Fonts load locally so the first screen no longer waits on a font served from abroad; the PDF icon shrank from 28KB to 7.1KB.
Fixes
- One material failing to parse no longer takes the whole graph down; failed turns are not shown as completed.
Deck templates, steadier parsing and richer imported chats
Features
- Deck templates: upload your own pptx and generation reuses its colours and fonts; when a template lacks a layout it degrades instead of failing.
- Steadier uploads and parsing: several materials parse at once, stuck jobs are reclaimed, and scanned PDFs are recognised in parallel shards with page-level progress and pages filled in.
- Drag to reorder the library (folders and materials alike) or drag materials into folders, and the order sticks.
- Imported conversations render as rich-text bubbles on the note page, with 13 display bugs fixed; code blocks are light background, dark text, with syntax highlighting.
- Save icons added for Zhipu Qingyan, Tencent Yuanbao and Grok.
Improvements
- Very long materials are stored in partitions instead of being truncated away.
- Answers follow the language you write in, not the interface language.
- Upload and preview progress bars show real progress and no longer fake a fill when the total is unknown.
Fixes
- Decks with a swapped template lost their layout; switching materials could corrupt the body; large files timed out while parsing.
Deck style families, editable export, sharing
Features
- Deck generation gained "style families": pick a family, then swap themes within it.
- Exported HTML copies can be edited in the browser and saved back to the original file (one authorisation on first save).
- Public sharing now covers every material type — conversations, agent sessions, notes, files, audio and video; data backs up to a ZIP.
- Archives expand on upload and backups restore in one click; mind maps run end to end from generation to a library material you can edit.
- Membership expiry reminders by email.
Improvements
- Exports follow the interface language; deck export matches what you see on screen (fonts, line height, layout).
- Deleting a material or folder no longer leaves stray files taking up space.
- Brand name spelling and CJK/Latin spacing unified across the site.
Fixes
- Share links occasionally 404ed; mind maps failed on long materials; exported SVGs would not open and nodes would not collapse.
English added, PayPal, and free credit
Features
- A bilingual site shipped (Simplified Chinese / English) with a language switcher.
- PayPal payments alongside WeChat, and members can top up extra credit.
- New users get bilingual welcome documents on their first visit.
Improvements
- Pricing set to ¥15 / ¥40 / ¥132, with a 14-day trial and ¥5 of credit; the member page shows credit and expiry.
- The workspace opens faster (editor and graph load on demand).
- Deck and PDF previews moved to pdf.js, fixing black and blank screens; scrollbars hide until you scroll.
- Uploads are capped at 128MB with an early warning when you pick a file.
Fixes
- Uploads over 2MB failed on mobile networks; home page copy and footer wording disagreed.
Native agent actions, and mobile
Features
- Native agent actions: ask in plain language to save an answer as a note or to organise files; tool calls get their own cards with per-tool icons.
- AI messages can be edited in place or regenerated; conversation history pages in, and a saved-note card survives when you look back.
- WeChat sign-in and sign-up shipped; @-mentioning a material no longer needs a leading space; microphone failures explain themselves and can be re-authorised from the icon.
- Account usage breaks down tokens by model and by day; external agents (an AI inside your editor, a command-line tool) can read and write your library, with install scripts for Windows, macOS and Linux.
- The library supports dragging materials into folders, nested folders and pinning; the workspace gained micro-interactions (pane collapse, message entry, button feedback).
Improvements
- Thinking traces render as Markdown with an animation while you wait; reply font size, block spacing and tool-card density were redone.
- Fixed a lag in the input box that re-rendered the whole chat pane on every keystroke.
Fixes
- Follow-up questions are no longer injected as a repeat of the previous turn; read-only views stopped showing a pointless notice.
The library, and editing in place
Features
- The library became one tree: folders and materials together, with nested folders, pinning (remembered on the server), tags and drag-to-file.
- Edit a material inside the conversation: per-block suggestions you can accept or reject, each with its reason; paired quote marks merge into one revision, halving the clicks.
- Materials and folders use model logos and file-type icons; folder rows gained a "···" menu and a context menu (new file, new subfolder, rename, pin, delete).
- A ⌘K command palette; ⌘⇧F / Ctrl⇧F focuses the search box.
- The home page was rewritten around three pillars with interactive demos (ask about a file, revise suggestions in place).
Improvements
- Very large materials open in milliseconds instead of seconds; list and detail split with skeletons.
- Pane corners and collapse animations tightened up; reply font size and spacing unified.
Fixes
- Pinned order was overwritten by the sort rule, the list flickered on refresh, and edit-suggestion bubbles could not be clicked.
Parsing, citations and search
Features
- Web link import: titles and bodies are fetched, images render, and the dialog closes itself when the import finishes.
- Parsing expanded a lot: Word, Excel, image OCR, PDF layout reconstruction and scanned pages; decks parse into structured HTML with slide previews.
- Cited answers: markers are clickable, ranges like [A-B] expand, and clicking jumps to the matching spot in the material.
- Search overhauled: live dropdown, bodies included, multiple snippets with highlighting and jump-to.
- The first layer of the knowledge network (automatic linking plus a full-screen graph), and quick capture notes.
Improvements
- Q&A moved to a new model engine, on by default, with an automatic fallback when it fails.
- Switching materials no longer blocks saving and shows loading feedback; search requests gained a timeout so they cannot spin forever.
Fixes
- Titles and bodies could get out of sync when switching materials; search could hang on "searching" indefinitely.
Materials, and a new workspace
Features
- Import PDF, Word, Excel, PowerPoint and web links.
- Selected passages join the conversation as context; paragraph-level AI suggestions can be accepted or rejected in place.
- Voice input; "/" shortcut menus in both the editor and the chat box.
Improvements
- A new AI chat pane: tool-call cards, copy and stats, with browsable history.
- Overall layout and interaction polish.
A new workspace skeleton
Features
- The workspace moved to a new three-pane skeleton: materials on the left, content in the middle, AI on the right.
- Uploaded materials parse automatically, and you can ask about a single selected material.
Improvements
- Note folder collapse state is remembered.
Fixes
- Occasional false "you are signed out" prompts.
v1 released
Features
- Notes gained nested folders; conversation titles can be renamed, and the layout is more flexible.
- Compare two notes line by line.
Improvements
- Markdown preview background set to white.
- Switched to the release version and went public (v1).
Filling in the site
Improvements
- The home page and its hero were redone.
- Company, business and product pages filled in; the footer tidied up.
- The site moved to a light theme; the company logo appears in the browser tab.
Accounts, and a rebuilt site
Features
- Password reset; the email field fills itself in after verification.
- Body font is configurable; rich text became the default editor.
Improvements
- The site was rebuilt around real product screenshots and a new home page layout.
- A welcome guide on the first visit.
Rich text, in full
Features
- The rich-text editor filled out: format painter, heading-level preview and quick settings, text colour, hyperlinks, font size.
- Invitations, and annotations on selected passages.
Improvements
- Privacy and about pages; a proper tab title and icon; payments went live.
Fixes
- The word cloud stopped rendering.
Membership, payments and a data panel
Features
- A member page (two columns), plan and order modules, QR payment and automatic navigation after paying.
- Order history and subscription expiry management; email verification on signup.
- A data panel: conversation and note counts, a word cloud, word counts, unread badges.
Improvements
- Saving a note is faster; feedback can be sent straight to our inbox.
Fixes
- Broken export and sharing, note dragging, search still filtering after clearing the box.
Rich text, and the first launch
Features
- Rich-text notes, and rich text became the default editor; columns move, items get context menus.
- Right-click a conversation to create a note from it.
- Qwen and Hunyuan added to collection.
Improvements
- Notes save automatically; the site went live publicly for the first time.
Page skeleton and dragging
Features
- Pages link to each other directly, with a shared top bar; the login page was rebuilt.
- Notes pin and drag to reorder; a link can open a specific conversation.
Improvements
- Lists show placeholder skeletons while loading; your conversation grouping preference is remembered.
Fixes
- Markdown rendering, and the note rename field.
Notes get their own column
Features
- Notes split out of the workspace (new, save, rename, menu) with search.
- Conversations can be grouped by criteria.
Improvements
- The divider between notes and workspace drags to resize; empty lists explain themselves.
Fixes
- Duplicate notes, content scrambling on refresh, messages lingering when switching conversations.
Six platforms, one library
Features
- Collection covers six platforms: DeepSeek, Doubao, Wenxin, OpenAI, Gemini and Kimi.
- Collected pages lay out as AI left, user right, preserving the original formatting; the note column splits into collected and notes.
Improvements
- Platform icons in lists and previews, with links back to the original chat.
The first version
Features
- The three-pane workspace took shape: list left, notes middle, chat right.
- The browser extension was built alongside the web app, uploading collected chats straight into your library.
Improvements
- Markdown preview and editing, click-and-drag text reorganisation, export and download, autosave.
- Search worked — being able to find things came first; sharing and sign-up shipped alongside.
Keep reading
CreateLLM-notebook: turn what you have read into a quiz you can practise
Having seen something is not the same as having learned it. LLM-notebook can turn saved articles, courses, product manuals or notes into a quiz: test yourself, review the answers and explanations, ask AI about mistakes, or export to PDF / Word.
Read postProductLLM-notebook: Support for bring-your-own model API Key (BYOK)
LLM-notebook now supports bring-your-own-key (BYOK) for DeepSeek, Qwen, OpenAI, Claude, Gemini, Kimi, GLM and MiniMax. Calls made with your key are billed by the provider and do not consume LLM-notebook token credits; LLM-notebook built-in models and BYOK models stay clearly separated in the selector.
Read postAskLLM-notebook: material you can question, answers you can trace
AI is changing how we learn from knowledge. In LLM-notebook, material is more than a static object you can read and search: it can be understood and questioned flexibly and personally, which greatly improves the efficiency of learning. On that basis, answers with citations make conclusions traceable, @-mentioning material organizes context across sources, and asking from a selection narrows inquiry to a single word. What emerges in the conversation can also settle into notes, so knowledge compounds.
Read post