Document Finder logo Document Finder
GitHub Download

Find research.
Packaged for AI.

Searches open-access sources, downloads the papers, and bundles them into a local library — ready to hand to any AI. No accounts, no API keys.

Download See how it works
Free & open source · AGPLv3 · macOS · Linux · Windows · No accounts, no telemetry

Sources

Eight open-access sources plus a built-in web meta-search, queried in parallel and deduplicated by DOI and content hash. No API keys.

arXiv arxiv.org

Preprints in CS, physics, math, biology. Direct PDF links, no rate-limit pain.

2.4Mpapers
OpenAlex openalex.org

~250M scholarly works with an open-access filter applied by default.

250Mworks
Semantic Scholar semanticscholar.org

~200M papers with semantic relevance ranking and citation graphs.

200Mpapers
Europe PMC europepmc.org

Life-science papers and preprints with open full text — biomedicine, genomics, clinical.

40M+abstracts
Internet Archive archive.org

Books, papers, scanned media. Slow but deep — perfect for older work.

40M+items
DOAJ doaj.org

Directory of Open Access Journals — vetted, peer-reviewed publications only.

20Kjournals
Zenodo zenodo.org

CERN’s open repository — papers, preprints, datasets and theses across every field.

3M+records
Gutenberg gutenberg.org

70,000+ public-domain ebooks. EPUB native, ideal for the history of ideas.

70Kebooks
Web meta-search built-in

Aggregates DuckDuckGo, Bing, Brave, Mojeek, Marginalia & Startpage behind a circuit breaker — plus a no-Docker local SearXNG fallback. Catches the long tail.

6+engines, zero setup

How it works

  1. 01

    Ask in plain English.

    Type what you’re looking for. Document Finder automatically splits your question into three or four scholarly sub-queries — and an optional local model sharpens them further. You always see exactly what it searched, so nothing is hidden.

    “reinforcement learning robotics”
    ↓
    01 reinforcement learning manipulation ●●●○○
    02 sim-to-real transfer robotics ●●○○○
    03 deep RL policy gradient robot ●●●●○
  2. 02

    Fan out across every source.

    Eight open-access sources plus web meta-search, queried in parallel, with per-source rate limiting and retries. You watch the progress bars fill, the source lanes climb, and the queue drain in real time. Stop any time.

    arXiv14
    OpenAlex9
    Semantic7
    EuPMC4
    IA3
    DOAJ5
    Zenodo2
    Gut.1
    Web4
  3. 03

    Save to a library.

    Each search gets its own folder on disk. PDFs and EPUBs land in the folder, extracted text in _text/, and a single library.db holds the metadata. Browse, search, open any document or reveal it in your file manager, export as ZIP, or feed straight into your favorite LLM tool.

    📁 reinforcement-learning/
    📄 attention-is-all-you-need-a1b2c3.pdf· 1.1 MB
    📁 _text/· 41 files · 18 MB
    📄 library.db· SQLite · 1.2 MB

Features

Parallel & resilient

Concurrent downloads with adaptive backoff and silent retries. A big run streams in while you watch, not after — then retry any failures in one click, and skip re-downloading documents you already have across libraries.

Local first

Your queries, your library, your machine. Optional on-device AI (a small embedding model plus a 1.5B LLM) reranks results and expands queries — fully offline, no API keys, no telemetry.

Full-text extraction

Every PDF and EPUB is run through a fast Rust extractor and stored alongside as plain text. Grep-able. LLM-ready. No vendor lock-in.

One-click export

Bundle any library into a portable .zip of PDFs, EPUBs and extracted text — ready to drop into any AI context window.

Live telemetry

Sub-query progress, per-source lanes, throughput, ETA, and a per-engine health bar for the web meta-search. For people who like watching their work.

Accessible & themeable

Three themes — Paper, Slate and Midnight — with nine accent colors and a compact/regular density. WCAG-AA contrast, full keyboard navigation, screen-reader live regions, and reduced-motion support throughout.

Specs

Stack
Tauri 2 Rust Solid.js SQLite
Footprint
Native — no Electron, no bundled browser
Local AI
Optional, downloaded on first use, run offline: BGE-small reranks results, Qwen 2.5 1.5B (~1.1 GB, Apache-2.0) expands and filters queries. No API keys, ever.
Sources
Eight open-access APIs (arXiv, OpenAlex, Semantic Scholar, Europe PMC, Internet Archive, DOAJ, Zenodo, Gutenberg) plus a built-in web meta-search over six engines. No keys, no quotas.
Storage
Plain files on disk — PDFs and EPUBs in the query folder, extracted text in _text/, metadata in a per-library library.db (SQLite).
Permissions
Filesystem (your library folder) and network (the sources you enable). No telemetry.
License
AGPLv3 — copyleft; source on GitHub, contributions welcome.
Made by
Web World Wide.

Download

Native installers for macOS, Windows, and Linux. Not notarized (macOS is ad-hoc signed) — first-launch steps are on the release page.

Download for macOS Apple Silicon · macOS 11+ .dmg
Windows.exe · .msi Linux.AppImage · .deb · .rpm · Flatpak
Not notarized · verify on the release page · source
Linux .deb / .rpm / .AppImage need glibc 2.39+ (Ubuntu 24.04+, Fedora 40+) — on older distros use the Flatpak, which bundles its own runtime.