01Why it exists
A pile, turned into a library
Manuals, alarm sheets, receipts, service notes, screenshots of things you meant to read later: they pile up in a Downloads folder where search means scrolling. DocCrawler turns that pile into a library you can search and question. The phone build runs in Termux with the internet off; the version below runs entirely in your browser.
02Parts list
- IngestEvery file is hashed with SHA-256 and de-duplicated. Add the same receipt twice and it's kept once.
- ExtractText straight from text and Markdown, pdf.js for PDFs, mammoth for Word, Tesseract OCR for photos of paper (each loaded only when needed).
- IndexRanked full-text search where a natural question is searched as any of its words, not a strict AND, so real questions find real answers.
- TagEach document gets its top keywords as tags; add, remove or filter by your own.
- The LibrarianAnswers grounded in ranked excerpts with numbered citations. Turn the model off and it answers with plain retrieval, and nothing leaves the device.
- YoursThe library lives in this browser's storage; export it as one file, import it anywhere.
03Construction
The hard part (phone build)
Ollama has no Android build, so on the phone it runs inside a Debian container under proot-distro. Keeping it up took a run of fixes: a pkill that killed its own wrapper, a server that died with its launcher, a wrong HOME that made every model vanish, and timeouts long enough for a phone CPU to finish thinking.
The browser build
IndexedDB for the library, BM25 over passages for search, pdf.js, mammoth and Tesseract.js on demand, and AutoFaber's free models for the Librarian when you allow it.
04Tolerances
Single-user on purpose. Auto-tags are a keyword heuristic, not a classifier; OCR is only as good as the photo. The Librarian cites its sources: check them for anything that matters.
what it is not, said plainly
05Inscription
The library stays on the device. The questions do too, if you want them to.
06The same room
- BallastEvery bill on a real calendar, ninety days projected forward, and one honest number: what's safe to spend today
- Sleep CompressionA sleep-restriction protocol turned tracker
- CryptoVaultEncryption you can read to the bottom of: RSA-4096 and AES-256-GCM in one HTML file, with identities, contacts, signatures and files
- Worm EaterAn ingredient scanner for food, cosmetics and personal care that says "I don't know" instead of quietly scoring the unknowns as safe
- OmnilogosRead any chapter of the Bible side by side in up to 22 public-domain texts, from the Hebrew, Greek and Latin originals to English and five other languages, and follow shared words and themes through a live graph.
- Mechanic's ArsenalMachines you can take apart in the browser