Ask questions about the books you own.
Every sentence cited from the text. Nothing past the chapter you're on.
Live beta · User guide · How it works · Blog · Issues
LORE is an open-source AI reader for your e-books. It turns each book into a knowledge graph of its characters, places, groups and events, then answers your questions from the text itself: every sentence in an answer cites the passage it comes from, and a second pass checks each one against that passage and cuts what the book doesn't say. Tell it how far you've read, and it never looks past your chapter.
- Answers you can check. Every sentence cites its passage; click a citation to read it in the book.
- Spoiler-safe. Set your reading progress per book. Answers, the graph, character profiles and timelines all stop at your chapter, enforced at retrieval, so the model never sees what comes later.
- Questions across a series. How a character changes from book one to book five, how a world or magic system works, who knows whom.
- Deep dives. Long, sectioned analyses with LORE's reading, the other side, and what the books leave open, all cited.
- Explore the graph. Click any character for a profile, a story timeline and their relationships, each with the passage behind it.
- Bring your own model. Gemini (the free tier works), Claude or OpenAI for reading and writing; embeddings run locally on Ollama.
- Use it from Claude. An MCP server lets claude.ai, Claude Desktop, Claude Code or any MCP client search your library and ask LORE questions.
- Private by design. Self-host it on your own machine, or use the hosted beta where each reader's library is kept separately.
The fastest way is the hosted beta at lore.jrvn.site: request an invite on the homepage. New accounts start with A Study in Scarlet already processed, so you can ask questions straight away. You'll need a free Gemini API key (or a Claude or OpenAI key); the app walks you through it.
With Docker, for you and the people you add:
curl -fsSLO https://raw.githubusercontent.com/Jatin-Roopchandani/LORE/main/docker-compose.yml
docker compose up -d
docker compose logs lore | grep "setup code" # a one-time code
# open http://localhost:8080, enter the code and create the admin accountChoose the AI model on the admin page: Gemini, Claude or OpenAI with your key, or a model you run yourself (llama.cpp's llama-server, vLLM, LM Studio, Ollama). Each person you add gets their own private library. The self-hosting guide covers requirements, a local model step by step, adding people, upgrades, backups, HTTPS and troubleshooting.
For development, or a single-user install with no Docker.
Prerequisites: Python 3.12+, uv, Node.js 18+ and Ollama.
# 1. The local embedding model (keep Ollama running)
ollama pull nomic-embed-text
# 2. A model key: a free Gemini key from https://aistudio.google.com/apikey
export GEMINI_API_KEY=your-key-here
# 3. Install, and check everything is in place
git clone https://github.com/Jatin-Roopchandani/LORE.git
cd LORE
uv sync
uv run lore doctor
# 4. Build the web app (once)
cd web && npm install && npm run build && cd ..
# 5. Start it
uv run lore uiOpen http://127.0.0.1:8766, click + Add book, and choose an EPUB or PDF you own. The Library panel shows processing as it goes (a novel takes a while on Gemini's free tier); when it's done, ask away.
Prefer the command line? lore ingest, lore process and lore ask do the
same; see the user guide. Instead of the environment variable,
the key can also live in ~/.config/lore/gemini.env (mode 600).
No cloud key? Run every step on your own model (llama-server, vLLM, LM Studio or Ollama): see Running fully local.
LORE runs an MCP server, so Claude can search your library and ask LORE questions with your reading position applied:
- Hosted beta: in claude.ai, open Settings → Connectors → Add custom connector and paste
https://lore.jrvn.site/mcp. - Self-hosted: run
uv run lore mcp(stdio) oruv run lore mcp --http.
Step by step: Use your ebook library in Claude and the user guide.
book (EPUB/PDF) → chapters & passages → LLM extraction → entity resolution
→ Kuzu graph + LanceDB vectors
question → split into parts → starting entities → graph walk → passages
→ claims, each quoting a passage → check every claim → cited prose
Reading progress becomes a per-book chapter cap applied as a prefilter on every vector search, graph query and profile, so later chapters are never retrieved. Behaviour changes ship behind switches, each judged by blind paired evals on held-out questions. The full pipeline, models and design decisions are in How LORE works.
Stack: Python, FastAPI, Kuzu (graph), LanceDB (vectors), SQLite and Postgres, Ollama (embeddings), React and TypeScript. The hosted beta runs on Docker Compose behind a Cloudflare Tunnel (deploy/).
LORE is in beta. It works end to end on full novels and series, and the test suite (4,000+ server tests, 800 web tests) runs on every push. Known limits: answers are only as good as the extraction, long series take a while to process, and a book must be one you can upload as EPUB or PDF.
Bug reports, ideas and pull requests are welcome. See CONTRIBUTING.md for setting up a development environment and running the tests. To report a security issue, please follow SECURITY.md rather than opening a public issue.
MIT © Jatin Roopchandani


