Cogito

Cogito, ergo sum. I think, therefore I am.

macOS PDF reader for active reading. Open a PDF, take edge notes beside each page, look up any word, and generate a NotebookLM video overview for any chapter. Chapter structure is detected automatically on-device with a local LLM when the PDF has no embedded outline. Reading progress is restored when you reopen a book.

Features

PDF reading with automatic margin cropping, single and two-page layouts, zoom, bookmarks, and full-text search. Reading progress is saved per book and restored on reopen.

Edge notes in two-page mode: narrow panels beside each page for free-form notes.

Word translation: select any word and get a Wikipedia-powered card in one of eight languages.

Chapter video overviews: click the video icon on any chapter's first page in two-page mode, or on the chapter row in the outline sidebar. Cogito extracts that chapter, uploads it to Google NotebookLM with an animation brief, and streams generation progress. The finished MP4 plays in a full-window overlay with captions. Chapters with existing videos show a green checkmark in the outline.

Ask Question (Cmd+J): type a question, and RAG retrieval finds the most relevant page instantly via BM25 keyword matching. The app navigates there, highlights the key term, and streams an ELI12-style answer from the on-device LLM with a closing metaphor in the reader's target language.

On-device LLM: Gemma 4 E4B via mlx-swift handles TOC detection, Ask Question answers, and mind map generation. Runs on the Apple Silicon Neural Engine, no internet required.

Architecture

%%{init:{'theme':'base','themeVariables':{'primaryColor':'#6366f1','primaryTextColor':'#1e1b4b','lineColor':'#94a3b8','fontSize':'13px'}}}%%
graph TD
    User(("User"))

    subgraph App[" Cogito.app "]
        direction TB

        subgraph Views[" SwiftUI Views "]
            direction LR
            CV["ContentView"]
            SB["SidebarView"]
            PR["PDFReaderView"]
        end

        VM[["PDFViewModel\nnav · bookmarks · outline · search\nnotes · video state · reading progress"]]

        PDFKit[["PDFKit"]]

        subgraph Svc[" Services "]
            direction LR
            LLM["LLMService\nGemma 4 E4B\non-device"]
            WTS["WikiTranslationService"]
            NLM["NotebookLMService\nProcess actor"]
        end
    end

    subgraph Ext[" External "]
        direction LR
        PY["generate_video.py\nnotebooklm-py"]
        Wiki[("Wikipedia API")]
        NLMAPI[("Google NotebookLM")]
    end

    User --> Views
    Views --> VM
    VM --> PDFKit

    VM -- "TOC pages" --> LLM
    LLM -- "OutlineNode[]" --> VM

    VM -- "selected word" --> WTS
    WTS --> Wiki
    Wiki --> WTS

    VM -- "chapter PDF" --> NLM
    NLM -- "spawn" --> PY
    PY --> NLMAPI
    NLMAPI -- "MP4" --> PY
    PY -- "JSON status" --> NLM

    style App fill:#fafafe,stroke:#6366f1,stroke-width:2px,color:#4338ca
    style Views fill:#eff6ff,stroke:#3b82f6,stroke-width:1.5px,color:#1e40af
    style Svc fill:#fdf2f8,stroke:#ec4899,stroke-width:1.5px,color:#9d174d
    style Ext fill:#f0fdf4,stroke:#4ade80,stroke-width:1.5px,color:#14532d
    style VM fill:#eef2ff,stroke:#6366f1,stroke-width:2px,color:#3730a3
    style PDFKit fill:#dbeafe,stroke:#60a5fa,color:#1e3a5f
    style LLM fill:#fce7f3,stroke:#f472b6,color:#831843
    style WTS fill:#fce7f3,stroke:#f472b6,color:#831843
    style NLM fill:#fce7f3,stroke:#f472b6,color:#831843
    style PY fill:#fffbeb,stroke:#f59e0b,color:#78350f
    style Wiki fill:#dcfce7,stroke:#4ade80,color:#14532d
    style NLMAPI fill:#dcfce7,stroke:#4ade80,color:#14532d

Loading

All views share one PDFViewModel. Services are injected as actors and communicate back via AsyncStream or @Published properties. The video path is the only one that crosses a process boundary: NotebookLMService spawns generate_video.py and reads JSON lines from stdout.

Project structure

cogito/
├── Sources/Cogito/
│   ├── CogitoApp.swift                  # entry point, menu commands
│   ├── ContentView.swift                # root layout, toolbar, video overlay
│   ├── PDFViewModel.swift               # all app state
│   ├── PDFReaderView.swift              # PDFKit bridge, word selection
│   ├── SidebarView.swift                # outline / thumbnails / bookmarks / videos
│   ├── CornellNoteView.swift            # edge note editor
│   ├── TranslationCardView.swift        # word translation card
│   ├── VideoGenerationBannerView.swift  # generation status banner
│   ├── NotebookLMService.swift          # process actor, status streaming
│   ├── LLMService.swift                 # mlx-swift / Gemma wrapper
│   └── WikiTranslationService.swift     # Wikipedia API client
│
├── Scripts/
│   └── generate_video.py    # uploads chapter PDF, polls NotebookLM, saves MP4
│
├── Package.swift            # mlx-swift via SPM
└── Makefile                 # build / bundle / run

Requirements

	Version
macOS	14+	SwiftUI, PDFKit, AVFoundation
Swift	6+
Python 3	any	video bridge
notebooklm-py	0.3+	NotebookLM client
mlx-swift	0.21+	on-device inference

Building

pip install notebooklm-py
notebooklm login          # one-time, browser-based Google auth

make build && make run    # dev build
make bundle               # full .app with mlx.metallib and generate_video.py

Video generation

Requires a Google account. Auth persists in a local cookie store via notebooklm-py.

generate_video.py receives a chapter PDF and the target output path from Swift, uploads the PDF plus an animation brief to a new NotebookLM notebook, and polls until the MP4 is ready. Status comes back as JSON lines on stdout; NotebookLMService parses them into AsyncStream<VideoStatus>.

Videos cache to ~/Library/Caches/com.cogito.app/Videos/ with a per-book hash in the filename.

License

MIT