Semantic search over a curated library of public-domain literature.

Live at askthecanon.com.

Instead of grepping for words, you bring a question — "how should I deal with people who wrong me?" — and get the passages that mean that, ranked across every book in your library and cited by author, title, and chapter. Matching is by meaning, not keywords, so a passage scores high even when it shares no words with your query.

Everything runs locally. Books come from Project Gutenberg via the Gutendex API; passages are embedded once with all-mpnet-base-v2 (no API key, no network at query time) and matched with plain NumPy cosine similarity.

1. Index (once per book): each book is split into ~250-word passages, tagged with their Book/Chapter heading, embedded, and cached to books/<id>.{txt,chunks.json,npy,meta.json}.
2. Ask (per query): only your question is embedded — one small vector — then matched against the cached library matrix. The corpus is never re-embedded.

The library is just library.txt: a list of Gutenberg IDs you grow over time.

The same core lives in main.py (chunk, embed, search) and is exposed two ways: the CLI above, and a thin FastAPI shell in web.py that serves a browser UI. The library matrix is loaded once and cached in memory; only the query is embedded per request. Searches and saved quotes are recorded to a local SQLite database (classics.db) via SQLModel.

uv sync

# find a book's Gutenberg id
uv run main.py search tolstoy war and peace

# download one book's text to books/
uv run main.py fetch 2600

# reconcile books/ with library.txt: build new ids, drop removed ones
uv run main.py sync

# ask a question across the whole library
uv run main.py ask "how do I face death without fear"
uv run main.py ask "the meaning of suffering" -k 8        # return 8 passages
uv run main.py ask "ivan meets the devil" --book 28054    # limit to one book

ask prints the top passages with citations, then lets you pick one to read in full.

uv run fastapi dev web.py    # or: just server

Open the printed URL and ask a question in plain words. Each result is a real passage with its citation. From there you can:

• Copy any passage to the clipboard.
• Pick several passages and Download PDF — a typeset A4 sheet of your question and the chosen excerpts, with the best-matching span underlined.
• Share as image — render a single passage as a shareable card.

Searches and saved quotes are recorded to classics.db (SQLite).

The repo ships a Justfile wrapping the common commands — run just to list them (just server, just ask "...", just reindex, just fetch <id>, just search ...).

This searches what philosophers and novelists actually wrote, so meet the corpus in its own register. It shines on timeless human questions — death, love, fear, meaning, ambition — asked in plain words. It struggles with modern jargon: a query like "how to be a better marketer?" scores low not because the books are silent on it, but because the word is anachronistic — Montaigne wrote about rhetoric and persuasion, Thoreau about honest trade, but never about "marketers". Ask in the books' own terms ("how do I persuade people and win them to my view?") and the same ideas surface with much higher scores.

You don't need to match the source text's words — the model handles paraphrase well. You need to match the theme. Questions that land well:

• how do I face death without fear?
• is it better to be feared or loved?
• is wealth worth pursuing?
• how do I forgive someone who wronged me?
• what makes a friendship last?
• how do I find courage when I am afraid?

Modern jargon names a role or activity the canon never knew. Reframe it as the human concern underneath, and the score roughly doubles:

Scores are relative: a top match around 0.45+ is strong. If even the best match falls below MIN_SCORE (0.34) the question is treated as off-domain and nothing is shown — rather than five weak matches dressed up as answers. Nonsense and modern-jargon queries top out around 0.28, so the floor catches them while leaving real questions (which clear 0.40) untouched.

library.txt is the source of truth. Add a Gutenberg id (use search to find it) or delete a line, then run uv run main.py sync: it builds embeddings for new ids, deletes the books/ files for removed ones, skips anything already built, and reports any id it can't fetch — so re-running is always safe. (It refuses to run if library.txt is empty, rather than wiping books/.)

uv run ruff format .
uv run ruff check --fix .
uv run ty check .
uv run pytest -q