How to ask AI questions about your own ebooks, with citations to the passage

Ask a general chatbot about a book and it will usually answer. The trouble is knowing where the answer came from. It may be from the book. It may be from a review the model read during training, a summary of a different edition, or nothing at all, stated with the same confidence. An answer you cannot trace to a page is a rumour about your book, not a reading of it.

The fix is old and simple: a citation you can follow. This post is about getting that from the ebooks you already own.

What a good answer looks like

Whatever tool you use, these are the things worth checking:

• It points at the passage. Not "chapter 3", but the sentences it relied on, one click away, so you can read them in context.
• It says when your books do not say. "Your library does not cover this" is a useful answer. A confident guess is not.
• It searches everything you own, not one file. The interesting questions cross books: which of them discuss an idea, and where they disagree.
• It keeps your notes apart from the text. What you wrote in the margin is your opinion; what the book says is the source. An answer should not blur the two.
• You know where your books go. Uploading a library is handing someone your reading history. Find out who stores it, and whether it trains their models.

The ways people do it today

Paste or upload into a chat assistant. This works for one book, or part of one. Long books strain what a model can take in at once, the answer can still lean on what the model already "knows", and a page number is only as good as the model's reading of it. Check the provider's settings for whether uploads are kept or used for training.

A notebook tool. Services such as Google's NotebookLM take a set of sources you upload and answer with links to the passages they used. They are good at it. They are built around a project, a notebook of sources for one piece of work, rather than a library you keep and read in for years.

Build your own. With a local model (llama.cpp or Ollama) and a tool such as AnythingLLM or Open WebUI, you can index your own files and ask them questions without anything leaving your machine. It is the most private route and the most work: you choose how books are split, which model embeds them, and how answers cite.

A library that answers. The last option is to ask in the place your books already live, which is what Lexarii is for.

How Lexarii answers from your books

Upload your EPUB and PDF files and ask a question, across your whole library or of one book you name. The answer:

• comes from your books, and cites the passages it used;
• links each citation to the page, opening the book at that line in the reader;
• says so when your books do not cover the question, rather than inventing an answer;
• keeps your notes apart from the author's words, when a note of yours bears on the question.

You can ask a follow-up, and the conversation carries on from there.

What it cannot do

Honesty about the edges, because they matter for research:

• Scanned PDFs with no text layer are shelved and readable as page images, but not searched: Lexarii does not do OCR.
• Books with DRM are not supported. Lexarii is for files you can open yourself.
• A model can still misread a passage. Citations are there so you can catch it. Treat an answer as a pointer to the text, and the text as the authority.

Where your books go

On lexarii.com, your books are stored in the EU, and searching them happens on our own servers. To write an answer, the model receives your question, the relevant passages, up to three of your notes and the conversation so far. It is reached through OpenRouter, in the United States, with the setting that asks providers not to store or train on your data. That is a request we make, not a guarantee we can give, and the privacy notice says how providers are chosen. Your books are never used to train models.

If that trade is wrong for you, Lexarii is also free software. Run it yourself with your own model, and nothing leaves your machine.

Try it

The demo is a shared library of public-domain classics, reset every hour: ask it about Pride and Prejudice or Frankenstein and follow the citations. When you want your own shelves, one plan covers unlimited books and questions.