DeepNote AI

Beta Release — This is the first public beta. Expect rough edges, evolving APIs, and occasional AI generation failures. We welcome bug reports and feedback via GitHub Issues.

A feature-rich, open-source desktop application inspired by Google's NotebookLM. Built with Electron, React, and powered by multi-provider AI (Gemini, Claude, OpenAI, Groq). Upload documents, chat with your sources, and generate studio-quality content — AI video overviews & music videos (Veo 3.1), AI podcasts, HTML presentation decks with a structured slide editor, image slide decks with drag-and-drop editing, PPTX export, whitepapers, infographics with video animation, flashcards, quizzes, mind maps, dashboards, literature reviews, competitive analyses, reports, and more — with agentic RAG, multi-agent generation pipelines, voice Q&A, cross-session memory, Knowledge Hub with file connectors and knowledge graph, and local embeddings.

• Features Overview
• What's New — Beta
• Tech Stack
• Getting Started
  • Quick Install (macOS)
  • Build from Source
  • Configuration
  • Running
  • Building
• Application Guide
  • Notebooks
  • Sources
    • Smart Source Recommendations
  • Chat
    • Interactive Artifacts
    • Artifact Shortcut Chips
    • Voice Q&A
    • Chat-to-Source Pipeline
  • Deep Research
  • Notes
  • Studio
    • Video Overview & Music Video
    • Audio Overview
    • Image Slide Deck
    • HTML Presentation
    • Infographic
    • White Paper
    • Study Flashcards
    • Quiz
    • Report
    • Mind Map
    • Data Table
  • Workspace (File Editor)
• AI Architecture
  • Agentic RAG
  • Multi-Agent Generation Pipeline
  • AI Output Validation Middleware
  • Cross-Session Memory
  • Tiered Embeddings
• Clipboard Quick-Capture
• Keyboard Shortcuts
• Project Structure
• Database Schema
• License

This beta includes 20+ major features transforming DeepNote AI from a notebook tool into a full intelligent research platform:

AI Models Used:

• gemini-3-flash-preview (default) / gemini-3-pro-preview / gemini-2.5-flash / gemini-2.5-pro — Chat, content generation, document analysis, agentic RAG
• claude-sonnet-4-6 / claude-opus-4-6 / claude-haiku-4-5 — Claude chat providers
• gpt-4o / gpt-4o-mini — OpenAI chat providers
• llama-3.3-70b-versatile — Groq chat provider
• gemini-2.5-flash-preview-tts — Multi-speaker text-to-speech (Kore & Puck voices)
• gemini-3-pro-image-preview (Nano Banana Pro) / gemini-3.1-flash-image-preview (Nano Banana 2) / gemini-2.5-flash-image (Nano Banana) — AI image generation for slides, infographics, whitepapers — selectable per creation
• veo-3.1-fast-generate-preview (default) / veo-3.1-generate-preview / veo-3-fast-generate / veo-3-generate / veo-2-generate — Video generation (image-to-video animation, 4-8s clips, up to 4K)
• text-embedding-004 — Text embeddings for RAG (Gemini cloud tier)
• all-MiniLM-L6-v2 — Local ONNX embeddings (offline tier, 384-dim)

1. Download the latest DMG from Releases
2. Open the DMG and drag DeepNote AI to your Applications folder
3. Launch the app, open Settings (gear icon), and enter your Gemini API Key
4. Click Test to verify, then Save — you're ready to go

Note: The app is signed but not notarized. On first launch macOS may show a security warning — right-click the app and choose "Open" to bypass it.

• Node.js 20+ and npm
• Google Gemini API Key (free at aistudio.google.com)

# Clone the repository
git clone https://github.com/Clemens865/DeepNote-AI.git
cd DeepNote-AI

# Install dependencies (ignore scripts to avoid premature native builds)
npm install --ignore-scripts

# Rebuild native modules for Electron
npx @electron/rebuild -f -w better-sqlite3

1. Launch the app
2. Click the Settings icon (gear) in the header
3. Enter your Gemini API Key
4. Click Test to verify, then Save

The API key is stored locally and never leaves your machine (except to call the Gemini API).

# Development mode with hot reload
npm run dev

# Preview production build
npm run start

# Full build (typecheck + bundle)
npm run build

# Platform-specific distributables
npm run build:mac      # macOS .dmg
npm run build:win      # Windows .exe
npm run build:linux    # Linux .AppImage

The Dashboard is the home screen. Each notebook is an isolated workspace with its own sources, chat history, notes, and generated content.

• Create a notebook by clicking the "+" button - choose a title and icon
• Workspace notebooks can be linked to a local folder for file editing
• Search notebooks by title using the search bar
• Delete notebooks via the context menu

Each notebook displays the number of sources and the last-modified date.

Sources are the knowledge base for your notebook. All AI features (chat, studio) draw from selected sources.

Supported Source Types:

Source Features:

• Select / Deselect - Toggle which sources provide AI context (checkbox per source)
• Source Guide - AI-generated summary overview of each source
• Auto-chunking - Documents are split into chunks with embeddings for RAG retrieval
• Delete sources individually

When viewing a source, DeepNote AI can recommend related sources from your other notebooks using vector similarity search. This helps you discover connections between different areas of research.

The Chat Panel is the main interaction mode. Ask questions and get AI-generated answers grounded in your selected sources.

• Streaming responses - Answers appear token-by-token in real time
• Source citations - Responses include [Source N] references linking to specific source chunks
• Suggested prompts - Quick-start buttons: "Summarize my sources", "Key takeaways", "Create a study guide"
• Upload from chat - Add new documents or audio files directly from the chat input
• Save to Note - Save any AI response as a note for later reference
• Clear history - Reset the conversation
• Cross-session memory - The AI remembers your preferences and learning patterns
• Knowledge Hub integration - When Knowledge Hub is configured with file connectors, the AI can search across all ingested knowledge. Results appear as context-enriched responses grounded in your knowledge base

Interactive Artifacts:

The AI can embed rich visualizations directly in its responses:

Artifact Shortcut Chips:

Six quick-action buttons appear above the chat input (when sources are selected). Click any chip to instantly request that artifact type - Table, Chart, Diagram, Kanban, KPIs, or Timeline.

Click the microphone button next to the chat input to start a voice conversation with your sources.

• Record - Tap the mic button to start recording your question
• Transcription - Audio is transcribed via Gemini AI
• RAG-powered response - Your question is answered using selected source context
• TTS playback - The AI's response is spoken back to you
• Transcript display - Full text transcript of the conversation appears in the overlay
• Multiple turns - Continue asking follow-up questions in the same session

Every AI response has action buttons to integrate it into your workflow:

Deep Research performs a multi-step, in-depth analysis of your sources. It goes beyond simple chat by running a longer AI pipeline with real-time progress updates.

• Click the Deep Research button in the chat panel
• Enter a research query
• Watch progress indicators as the AI analyzes sources in depth
• Results appear as a comprehensive research report in the chat

The Notes Panel provides a scratchpad for your own thoughts alongside AI-generated content.

• Create notes with the "+" button
• Auto-save - Content saves automatically as you type (500ms debounce)
• Editable titles - Click any note title to rename it
• Convert to Source - Turn a note into a source document so the AI can reference it in chat and studio
• Timestamps - Each note shows when it was last edited
• Delete notes with confirmation

The Studio Panel contains 17 AI-powered content generation tools. Each tool transforms your selected sources into a different output format.

All studio tools support:

• Custom instructions - Guide the AI's focus and audience
• Length options - Short, Default, or Long output
• Rename generated content
• Delete generated content
• Generation history - All past outputs are accessible
• Search & filter - Find generated items by title, filter by type, sort by newest/oldest/title/type
• Multi-agent pipeline - Complex types (report, whitepaper, literature review) use a Research → Write → Review pipeline with real-time progress indicators

Generate AI-powered narrated videos or music videos from your notebook sources.

Two Modes:

Two-Phase Pipeline:

1. Storyboard (Phase 1) — AI plans scenes → generates images (4 parallel) → presents storyboard for review
2. Animation (Phase 2) — After approval, generates narration (TTS) → animates scenes (Veo) → assembles final MP4 (ffmpeg)

Storyboard Review:

• Visual grid of all scene thumbnails
• Per-scene regeneration with custom instructions
• Approve and proceed to animation, or regenerate individual scenes

5 Veo Video Models:

Video Features:

• Narrative styles — Explain, Present, Storytell, Documentary
• Narration sync — TTS audio fitted to match each video clip duration
• Mood control — Auto, custom prompt, or reference image style
• Style presets — 6 visual presets + custom style builder with color picker
• Quota-aware — Auto-fallback between Veo model variants on 429; Ken Burns fallback on full exhaustion
• Music video auto-detection — Detects audio file duration via ffprobe, generates exact scene count
• Resolution selector — 720p, 1080p, 4K (filtered by model capability)
• Progress tracking — Per-scene progress with retry status and fallback indicators

Generate a podcast-style audio conversation between two AI speakers discussing your source material.

Format Options:

Length Options:

• Short: 8-12 conversational turns
• Default: 15-25 turns
• Long: 30-45 turns

Audio Features:

• Two distinct AI voices (Kore & Puck) via Gemini TTS
• Built-in audio player with play/pause/seek
• Full transcript view with color-coded speakers
• Download as WAV file

The most advanced studio feature. Generates complete slide presentations with AI-created background images and editable text overlays.

6 Visual Style Presets:

Custom Style - Upload a reference image and the AI will extract and replicate its visual style.

Image Model Selection:

Each generation lets you choose between three Gemini image models:

Presentation Formats:

• Presentation - Educational / informational slides
• Pitch Deck - Business pitch (Problem, Solution, Market, Ask)
• Report Deck - Data-driven analysis with charts

Slide Count: Test (3), Short (5), Default (10), Recommended (AI-optimized, up to 20)

Aspect Ratios: 16:9 (widescreen), 4:3 (classic), 1:1 (square), 3:4 (portrait), 9:16 (tall portrait)

Two Render Modes:

Rich Slide Editor (Hybrid Mode):

Click the pencil icon on any hybrid slide to enter edit mode:

• Drag & Drop - Freely position title, bullet, and text elements anywhere on the slide
• Rich Text Formatting - Bold, italic, underline, strikethrough via Tiptap editor
• Bullet Lists - Toggle bullet list formatting
• Text Alignment - Left, center, right alignment per element
• Font Size - Dropdown with sizes from 12px to 48px
• Text Color - 10-color palette picker
• Links - Add hyperlinks to text
• Add Text Box - Create new free-positioned text elements
• Delete Elements - Remove unwanted text boxes (minimum 1 must remain)
• Resize - Drag the corner handle to adjust element width
• Auto-save - All changes persist automatically (800ms debounce)
• Save Button - Explicit save with visual feedback (saving spinner → "Saved" checkmark)
• Click outside to deselect elements

Presentation Features:

• Fullscreen Mode - Press the expand button for a full-screen presentation
• Keyboard Navigation - Arrow keys and Space to advance slides (disabled while editing text to preserve cursor position)
• Thumbnail Strip - Visual slide navigator
• Speaker Notes - Show/hide AI-generated speaker notes per slide
• Download - Save individual slides as PNG files

Generate structured HTML presentations with a full slide editor. This is the newest presentation format, offering fine-grained per-slide editing with AI assistance.

7 Slide Layouts:

Slide Editor Features:

• Preview / Edit toggle - Switch between live HTML preview and per-slide editor
• Slide Navigator - Thumbnail sidebar with drag-and-drop reordering, add/delete slides
• Per-slide editing - Edit title, subtitle, speaker notes, and individual content blocks
• Layout picker - Change slide layout type on the fly
• AI Regeneration - Regenerate any slide with custom instructions
• Image Generation - Add image blocks and generate AI images via Nano Banana models
• Live Preview - Instant HTML preview refresh on every edit
• PPTX Export - Export to PowerPoint with embedded images, themed colors, and template support
• Prompt Templates - Save and reuse custom generation prompts

Generate AI-powered single-page visuals from your sources with format-specific design intelligence.

3 Format Types:

Aspect Ratios: 16:9, 4:3, 1:1, 3:4 (portrait), 9:16 (tall portrait — ideal for social stories)

Render Modes:

• Full Image - Visual-first: all text and graphics integrated into a single AI-generated image
• Hybrid - AI background with structured HTML text annotations overlaid

Style Options: 6 visual presets + custom style builder (upload a reference image for AI style matching)

Additional Features:

• Image Model Picker - Choose between Nano Banana Pro, Nano Banana 2, or Nano Banana
• Veo Animation - Generate 4-8s video clips from infographics via Google Veo 3.1
• Custom Color Palettes - Extract or define color schemes for consistent branding
• Download - Save as PNG image

Generate multi-section academic, business, or technical papers with AI-generated illustrations.

• Cover image - AI-generated cover art matching the paper's theme
• Section illustrations - Each major section includes a generated image
• Table of Contents - Auto-generated with clickable navigation
• References - AI-generated reference list based on source material
• 3 image models - Choose Nano Banana Pro, 2, or Classic per generation
• A4 PDF Export - Professional document layout with images, typography, and formatting

Generate interactive flashcards from your sources for study and review.

• Card Count: Fewer (5-8), Standard (10-20), More (25-35)
• Difficulty: Easy, Medium, Hard
• Interactive - Click cards to flip and reveal the answer
• Source References - Each card cites the relevant source

Generate multiple-choice quizzes to test your understanding.

• Question Count: Fewer (3-5), Standard (5-10), More (12-20)
• Difficulty: Easy, Medium, Hard
• 4 options per question
• Interactive answer reveal - Green for correct, red for incorrect
• Explanations - Each answer includes a "why this is correct" explanation

Generate a structured written report from your sources.

• Executive Summary + 3-6 detailed sections
• Custom focus - Use instructions to target specific topics or audiences
• Multi-agent pipeline - Research → Write → Review for higher quality
• Copy to clipboard - One-click copy of the full report text

Generate a hierarchical mind map visualization of your source content.

• 3 levels deep - Central topic, main branches, sub-branches
• 3-6 main branches with 2-4 children each
• Color-coded depth - Indigo (root), Blue (L1), Emerald (L2), Amber (L3)
• Expandable tree - Click to expand/collapse branches

Extract structured data from your sources into a sortable table.

• 5-15 rows of meaningful data points
• AI-optimized columns - The AI determines the best column structure
• HTML table display with clean formatting

Link a local folder to your notebook and browse, edit, and AI-index files without leaving the app.

File Tree:

• Hierarchical browser - Navigate your project's file structure
• .gitignore support - Respects .gitignore patterns for file exclusion
• Status indicators:
  • Gray - Unindexed (not included in AI context)
  • Green - Indexed (included in AI context)
  • Yellow - Stale (file changed on disk since last index)
  • Red - Error during indexing
• Select / Deselect files - Control which files become AI sources

File Editor:

• Text editing - Edit any text file directly in the app
• Dirty state indicator - Orange dot when there are unsaved changes
• Auto-detect read-only - Binary/unsupported files open in read-only mode

AI Rewrite:

• Select text in the editor and press Cmd/Ctrl + K
• An inline popup appears near your selection
• Type an instruction (e.g., "Make it more concise", "Add error handling")
• The AI rewrites just the selected text in context of the full file
• One-click replace of the original selection

Workspace Sync:

• Detects file changes (added, modified, deleted) on disk
• Re-indexes stale files to keep AI context up to date
• Shows sync status banner with counts

DeepNote AI uses several advanced AI patterns under the hood to deliver high-quality results.

Traditional RAG uses a single query to search for relevant context. DeepNote AI's Agentic RAG system reasons about your question first:

1. Sub-query generation - The AI generates 2-3 targeted search queries from your question
2. Multi-query search - Each sub-query is embedded and searched independently
3. Deduplication & ranking - Results are merged by chunk ID and ranked by combined score
4. Sufficiency check - The AI evaluates whether enough context was retrieved, and can request one additional retrieval round

This produces significantly better context for complex, multi-faceted questions.

For complex studio content types (report, literature review, competitive analysis, dashboard, whitepaper), generation uses a three-stage pipeline:

Real-time progress updates show which stage is currently executing.

All AI-generated structured content passes through a validation middleware pipeline:

• JSON validation - Strips markdown fences, validates JSON parsing
• Structure validation - Per-type schema checks (quiz must have questions with options, flashcards must have front/back, etc.)
• Retry with feedback - On validation failure, the error message is appended to the prompt and the AI retries (up to 3 attempts)

The AI learns from your conversations and remembers preferences across sessions:

• Automatic extraction - After each conversation, the AI extracts preferences, learning patterns, and context
• Memory types - Preference, learning, context, and feedback memories
• Confidence scoring - Each memory has a confidence score (0.0-1.0) that updates over time
• Per-notebook + global - Memories can be scoped to a specific notebook or applied globally
• System prompt injection - Relevant memories are automatically included in the AI's context

DeepNote AI supports a tiered embedding system for flexibility between online and offline use:

The system automatically selects the best available tier. Configurable via Settings (auto, gemini, or local).

A system tray icon provides quick access to capture clipboard content without opening the main window.

• Global shortcut - Press Cmd+Shift+N from any app to capture clipboard text
• Clipboard history - Last 10 captured items stored in the tray menu
• Add to notebook - Send any captured text directly as a source to your current notebook
• Always accessible - Works even when the main window is minimized

src/
  shared/
    types/              # Shared TypeScript interfaces & IPC channel definitions
      index.ts          # Core types (Notebook, Source, Note, ChatMessage, UserMemory, etc.)
      ipc.ts            # IPC channel names & handler type map
  main/                 # Electron main process
    index.ts            # App entry, window creation, protocol registration, tray init
    db/
      index.ts          # SQLite database initialization (auto-create tables)
      schema.ts         # Drizzle ORM schema definitions (including user_memory)
    ipc/
      chat.ts           # Chat message handlers (streaming, agentic RAG, DeepBrain context)
      knowledge.ts      # Knowledge Hub handlers (file connectors, knowledge graph, search)
      clipboard.ts      # Clipboard history & add-to-notebook handlers
      config.ts         # API key management
      memory.ts         # Cross-session memory CRUD handlers
      notebooks.ts      # CRUD for notebooks
      notes.ts          # CRUD for notes
      research.ts       # Deep research pipeline
      sources.ts        # Source ingestion, parsing, embedding, recommendations
      studio.ts         # Content generation & image slides (with pipeline progress)
      voice.ts          # Voice Q&A session handlers
      workspace.ts      # File tree, editing, syncing
    services/
      ai.ts             # Gemini AI (chat, generation, planning, memory-aware prompts)
      agenticRag.ts     # Multi-query agentic RAG system
      aiMiddleware.ts   # AI output validation & retry middleware pipeline
      chunker.ts        # Document chunking with token counts
      config.ts         # Persistent config store (API key, embeddings model)
      documentParser.ts # PDF, DOCX, TXT/MD parsing
      embeddings.ts     # Gemini embedding generation
      generationPipeline.ts  # Multi-agent Research → Write → Review pipeline
      htmlRenderer.ts   # HTML presentation rendering with image embedding
      imagen.ts         # AI image generation for slides
      knowledgeIngestion.ts  # Knowledge ingestion pipeline
      knowledgeStore.ts      # Knowledge storage and retrieval
      localEmbeddings.ts     # ONNX local embedding inference
      memory.ts         # Cross-session memory service
      rag.ts            # Retrieval-Augmented Generation (with agentic option)
      recommendations.ts     # Cross-notebook source recommendations
      sourceIngestion.ts     # Shared source processing pipeline
      tieredEmbeddings.ts    # Tiered ONNX → Gemini → hash embedding system
      tray.ts           # System tray icon & clipboard quick-capture
      pptxRenderer.ts   # PPTX generation with embedded images
      pptxTemplateParser.ts  # PPTX template parsing for theme extraction
      tokenTracker.ts   # API cost tracking (LLM tokens, Veo per-second, image gen)
      tts.ts            # Multi-speaker text-to-speech
      ffmpegAssembler.ts # Video assembly, Ken Burns fallback, narration sync
      veo.ts            # Veo video generation (5 models, quota-aware fallback)
      videoOverview.ts  # Video Overview pipeline orchestrator (storyboard + animation)
      vectorStore.ts    # In-memory vector similarity search
      voiceSession.ts   # Voice Q&A session management
      webScraper.ts     # URL content extraction
      workspace.ts      # Filesystem scanning & .gitignore
  preload/
    index.ts            # Context bridge API (typed IPC methods)
  renderer/
    src/
      App.tsx           # Router & top-level routes
      main.tsx          # React entry point
      styles/
        globals.css     # Tailwind v4 config, theme tokens, Tiptap styles
      stores/
        appStore.ts     # Global app state (theme, settings modal)
        notebookStore.ts# Active notebook state
        workspaceStore.ts# Workspace file tree & editor state
      hooks/
        useNotebooks.ts # Notebook listing hook
      components/
        chat/           # ChatPanel, ChatInput, ChatMessage, ChatKnowledgeResults, VoiceOverlay
        common/         # Button, Modal, Spinner, Toast, SettingsModal
        dashboard/      # Dashboard, NotebookCard
        layout/         # AppLayout, Header, ResizablePanel
        notes/          # NotesPanel, NoteEditor
        sources/        # SourcesPanel, SourceList, AddSourceModal
        knowledge/      # KnowledgeHub, ConnectorsTab, KnowledgeGraphTab
        studio/         # StudioPanel, ToolGrid, GeneratedContentView,
                        # ImageSlidesView, ImageSlidesWizard,
                        # HtmlPresentationView, PresentationSlideEditor,
                        # PresentationSlideNav, HtmlPresentationWizard,
                        # DraggableTextElement, SlideEditorToolbar,
                        # StudioCustomizeDialog
        workspace/      # WorkspaceLayout, FileTreeView, FileTreeNode,
                        # FileEditor, WorkspaceSyncBanner

DeepNote AI uses SQLite with 10+ tables:

All data is stored locally in ~/.config/deepnote-ai/ (or platform equivalent). No data is sent to external servers except Gemini API calls for AI features.

This is a beta release. The following issues are known and tracked for future releases:

We welcome bug reports at GitHub Issues.

This project is provided as-is for educational and personal use.