reader 3

A lightweight, self-hosted EPUB reader for reading one chapter at a time and optionally generating local AI summaries.

This repository is a fork of Andrej Karpathy's original "reader 3" demo — adapted to add a local, privacy-first summarization feature that runs on your machine using Ollama and a lightweight Qwen model so you can read and summarize on the go.

Why this fork exists

• Forked from: Andrej Karpathy's original project, kept simple and minimal so it remains easy to run and adapt.
• Local-first summarization: I chose Ollama to run models locally (no cloud keys), and qwen3:4b as the default model because it's small enough for many people to run on personal machines while providing useful summaries.
• Reading on the go: The app is intentionally minimal—run it on a laptop or small server and access it from your phone or tablet on the same network.

What I changed in this fork

• Added a summarization button in the reader UI (templates/reader.html) that requests page summaries.
• Implemented summarizer.py which uses a LangGraph workflow and calls Ollama to run qwen3:4b.
• Hooked the summarizer into the backend via the FastAPI endpoint /api/summarize/{book_id}/{chapter_index} in server.py.

Install and run Ollama, and pull the model used by this fork:

# Start Ollama server (runs on http://localhost:11434)
ollama serve

# Pull the lightweight model many can run locally
ollama pull qwen3:4b

Original Book Usage

The project uses uv. So for example, download Dracula EPUB3 to this directory as dracula.epub, then:

uv run reader3.py dracula.epub

This creates the directory dracula_data, which registers the book to your local library. We can then run the server:

uv run server.py

And visit localhost:8123 to see your current Library. You can easily add more books, or delete them from your library by deleting the folder. It's not supposed to be complicated or complex.

Contributing / Running Locally

• This fork is intentionally small and experimental. Feel free to modify the summarizer prompt in summarizer.py or swap models.
• If you improve things and want to share, open a PR against this fork.

License

MIT