Back to blog
Agent
Product
Create
Changelog
Everything we changed, from the first version on main to today. Versions count up in release order (v0.0 to v1), and every entry maps to a real commit.
Knowledge network, editable decks
Features
- Entity-relation extraction for the knowledge graph was rewritten and now runs block by block in parallel: a 20k-character material went from 31s to 9s.
- Revise a single slide with one sentence — only the current page changes, every other page is preserved word for word. Hand editing decks shipped in the same batch.
- New "LLM chat" tab in the sidebar, and a "+" button in the chat input that imports material directly.
Improvements
- Deck generation moved to ppt-master with per-page parallel rendering: 8 pages went from 649s to about 200s, with the same quality gate.
- Cross-material linking now asks per material and can backfill a folder, or everything, in one go.
- Fonts are self-hosted to remove a cross-border render block, and brand entity declarations now cover spaced spellings of the name.
Fixes
- When a linking result is cut off, complete edges are salvaged instead of discarding the whole batch.
Conversations and notes, polished
Features
- Drag to reorder the material library (folders and materials alike), and the order persists; drop targets tell folder and empty space apart.
- Imported conversations render as rich-text bubbles on the note page.
Improvements
- Added save icons for Zhipu Qingyan, Tencent Yuanbao and Grok.
Fixes
- Code blocks were invisible because text and background matched; now light on dark with multi-language highlighting.
- Switching material could corrupt the body text — the read-only rule is enforced in SQL now.
- The progress bar no longer draws a fake fill when the total is unknown.
Deck generation
Features
- Deck generation, direct editing, semantic edits, export and present mode (Phase A-E), with theme selection and a design system.
- Member data can be exported as a backup ZIP.
Improvements
- A local dev instance (dev:local on :3001) plus password-free dev sign-in, so development no longer occupies the production port.
- Material list queries moved to an explicit field whitelist, so body fields no longer leak into lists.
Launch week
Features
- Bilingual site: Simplified Chinese and English.
- v2 cut over for real: the whole site runs on our own machine (self-hosted behind an encrypted tunnel), with the cloud only forwarding traffic.
Improvements
- Edit proposals and save state are isolated per material, and concurrent PATCH requests are serialised instead of overwriting each other.
Fixes
- Twelve batches of fixes from the review list: material library, three-pane wiring, chat pane, inline editing, races and retry storms.
Polish and proofreading
Features
- Edit proposals carry a reason bubble per change, and no longer live in the chat pane.
- Proofreading rules integrated: paired quotes merge into one revision unit, halving the clicks.
Improvements
- Output discipline: model thinking is compressed, and answers follow the language the user writes in.
- Homepage copy was cut down to three features, with the sections below aligned to match.
New agent engine
Features
- The agent backend moved to DeepSeek Harness (M2-M4), on by default with graceful fallback; reply rendering, tool cards, copy and stats were rebuilt on the new engine.
Improvements
- Material library and import polish: URL import states (fetching to ready or failed), tighter card interactions, clipboard images in the import dialog.
- Search upgrade: covers bodies and conversations, live dropdown, keyword highlighting and jump-to.
Editing and parsing
Features
- Selected passages join the conversation as structured context, and paragraph-level AI suggestions can be accepted or rejected inline.
- "/" shortcuts in both the editor and the chat input, plus voice input built on the browser API.
- First layer of the knowledge network: related materials link up automatically, with a full-screen graph view.
Improvements
- PDF layout rebuilt as flowing paragraphs, and OCR for scanned and mixed PDFs.
- Answers come with paper-style citations that jump back to the source passage.
v2 foundations
Features
- The three-pane shell was rebuilt, with a feature flag for gradual rollout.
- Material layer MVP: the materials table with CRUD, an async parse job queue, and a standalone worker consuming it.
- AI layer: material summaries, Q&A over the selected material, and the first agent integration.
Improvements
- API split into route / auth / validation / service layers with unified responses and validation, across all 17 endpoints.
Website and v1 release
Features
- The homepage was rebuilt with real product imagery and a new hero layout.
- Account flows completed: forgot password, and the verified email auto-fills the next form.
Improvements
- Business, product and footer pages filled in, and the brand icon shows in the browser tab.
- v1 switched to a release build and went out publicly.
Conversations, notes, first revenue
Features
- "Create note from" on a conversation: one chat becomes one note.
- Payment page and orders, with subscription expiry management.
- Email verification added to sign-up.
Improvements
- Conversations can be refreshed, and the top-left logo was redone.
- Stats panel: conversation and note counts, a word cloud, and an unread badge.
Fixes
- A scrolling problem on the note page.
The first version on main
Features
- The three-pane workspace took shape: list on the left, note in the middle, chat on the right.
- Supabase integration moved the data off a local machine.
- The browser extension was built alongside the web app, so saved content could be uploaded straight into it.
Improvements
- Markdown preview and editing, click-and-copy text restructuring, export and file download.
- Search and remote search first: being able to find things is the premise for everything after.
Keep reading
AgentThree new shapes for a note, once the Agent lives inside
Note tools assume three jobs belong to you: writing, filing, carrying. With the Agent living inside, all three can be handed over — and the note itself gains new shapes.
Read postProductWhy we did not build a nicer chat box
The chat box is the default shape of every AI product, and also its ceiling: context is used once and thrown away. Notes on the trade-off we made.
Read postCreateFrom material to a deck you will actually use
Generating slides is easy. Wanting to use them afterwards is hard. Most of our effort went into the second half: editable, exportable, structurally sound.
Read post