DocCrawler

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, without sending a single file anywhere.

• Ingest — scan a folder or upload from the browser. Every file is hashed (SHA-256), deduplicated, and filed into library/YYYY/MM/.
• Extract — text straight from .txt/.md, pypdf for PDFs, python-docx for Word, Tesseract OCR for images. No OCR installed? The image is still kept, with a warning instead of a crash.
• Index — plain SQLite with an FTS5 full-text table: no server process, fast on a phone. Natural-language questions are searched as any of their terms, not a strict AND, so real questions find real answers.
• Tag — each document gets its top keywords as tags automatically; add, rename or remove your own.
• The Librarian — a multi-turn chat that always grounds itself in ranked excerpts with filename citations. It answers with a local model through Ollama first, Claude if a key is set, and plain retrieval if neither — degrading, never hanging.

Ollama has no Android build, so on the phone it runs inside a Debian container under proot-distro. Getting that to stay up took a run of fixes: a pkill that matched — and killed — its own wrapper script, a server that died the moment its launcher exited, a wrong HOME that made every model disappear, and chat timeouts long enough for a phone CPU to finish thinking. The setup and start scripts now carry all of it, and they are safe to re-run.

A single-user tool: it binds to 127.0.0.1 with no login, on purpose. Auto-tags are a keyword-frequency heuristic, not a classifier. OCR is only as good as the installed Tesseract, and a small local model on a phone is usable, not fast. The Librarian cites its sources — check them for anything that matters.