mrsibe/KnowNote

KnowNote – a local‑first research workspace

What it is

• A cross‑platform desktop app (Electron) that lets you turn your own PDFs, Word docs, PowerPoints and web pages into a searchable knowledge base.
• You can ask natural‑language questions and get answers that are cited back to the exact passage they were derived from.
• All heavy lifting – parsing, chunking, embedding and vector search – runs on your machine. The only network call is the request you send to the LLM endpoint you configure.

How it works

1. Import a document. The app uses pdfjs-dist, mammoth, officeparser and turndown to extract text while preserving structure (pages, headings, slides).
2. Split the text into ~500‑character passages with a small overlap; each passage records its start/end offsets for precise citations.
3. Embed each passage with the open‑source Xenova/multilingual‑e5‑small model (384‑dim, quantised to q8) via ONNX, running in the Electron main process.
4. Store embeddings in a local SQLite database (using sqlite‑vec via Drizzle ORM). Each notebook has its own vector table, so different embedding spaces never get mixed.
5. Query: you type a question; the app retrieves the most relevant passages, sends them (plus your prompt) to the chat model you configured (OpenAI, Anthropic, Google, Ollama, etc.), and displays the answer together with clickable citations.
6. Take notes: you can write structured notes beside the answer and turn the notebook into a mind‑map.

Key features

• Local RAG – no server, no Docker, no cloud storage; everything stays on your device.
• Bring‑your‑own‑model – any OpenAI‑compatible, Anthropic‑compatible, Google‑compatible, or local Ollama endpoint; the app does not ship a chat model.
• Traceable answers – every response lists the source passages and their exact location in the original document.
• Multi‑format import – PDF, Word, PowerPoint, and arbitrary web pages.
• Notes & mind maps – capture conclusions and visualise relationships.
• Open source – GPL‑3.0 license; the embedder model is downloaded once and cached.

What’s missing (as of the latest release)

• Audio upload/transcription, quiz generation and slide generation are still in development.
• No Linux build yet; macOS build is Apple Silicon only.
• Builds are unsigned (macOS) and not notarised, so you’ll see a one‑time security prompt.
• No multi‑user sync, collaboration, or telemetry – it is a single‑user tool.

Tech stack

• Electron + React + TypeScript + TailwindCSS – UI and desktop packaging.
• Tiptap – rich‑text note editor.
• ONNX runtime – runs the multilingual‑e5‑small embedder locally.
• SQLite + sqlite‑vec – vector storage and similarity search.
• Node.js – development environment; npm run dev launches the app in hot‑reload mode.

Getting started

1. Download the appropriate installer from the Releases page (.exe for Windows, ‑arm64.dmg for Apple Silicon macOS).
2. On macOS, clear the quarantine flag (sudo xattr -rd com.apple.quarantine /Applications/KnowNote.app).
3. On Windows, click Run anyway if SmartScreen warns you.
4. Open Settings → Models and add an endpoint (e.g., https://api.openai.com/v1/chat/completions with your API key, or a local Ollama server).
5. Import a document, ask a question, and explore the cited answer.

Typical use cases

• Academic research: ask questions over a collection of papers and see the exact sentence that supports each answer.
• Business knowledge management: query internal PDFs, slide decks, and policy docs without uploading them to a cloud service.
• Personal study: build a private “second brain” that you can interrogate with any LLM you trust.

License

• GPL‑3.0 – you can freely use, modify, and redistribute the code, provided you keep the same license.

Where to learn more

• Project website: https://knownote.pages.dev (detailed feature guide, design rationale, and comparison with NotebookLM/AnythingLLM).
• Architecture overview and design notes are in the repo’s DESIGN.md.
• Contribution guidelines are in CONTRIBUTING.md.

---

KnowNote is a genuine, open‑source AI‑augmented desktop tool that brings local Retrieval‑Augmented Generation to everyday research workflows while keeping data private and under user control.