How to Build a Local AI-Powered Document Search Engine with DocuBrowser

Table of Contents

Toggle

It Started With a Folder Called “Research”

I have this folder. Maybe you do too. It’s called something innocuous like “Research” or “Archive” — and inside it, there are PDFs I downloaded three years ago, Markdown notes from a conference I attended, Word docs that hold the only copy of an important report, and ebooks I keep telling myself I’ll read someday.

The problem isn’t storage. It’s findability. I know that somewhere in that folder is the paper about zero-trust architecture that a colleague asked about last week. But digging through hundreds of files to find it? That’s not a productive use of anyone’s afternoon.

I tried everything. Tagging systems. Spreadsheets. grep. Nothing stuck — because the real challenge isn’t searching for words I remember. It’s finding documents when I only have a vague idea of what they contain. That contract about the lease renewal. That paper with the network diagram I liked. You know the feeling.

Enter DocuBrowser — an open-source, local-only document search engine that combines traditional keyword search with AI-powered semantic understanding. And it runs entirely on your machine. No cloud, no API keys, no monthly subscription.

What Is DocuBrowser?

DocuBrowser (specifically DocuBrowse v0.9.0) is a Python-based web application that indexes your documents and lets you search them in two ways: old-school keyword matching and modern AI semantic search. Think of it as Google for your personal document collection — one that doesn’t phone home.

Under the hood, it uses SQLite FTS5 for fast full-text keyword search and Ollama with nomic-embed-text for semantic embeddings. It generates on-demand AI synopses using the dolphin3 model running locally. If you haven’t set up Ollama yet, I wrote a complete guide to running AI models locally that covers installation and model management.

DocuBrowser supports a wide range of formats: PDFs, DOCX, PPTX, XLSX, EPUB, MOBI, AZW3, HTML, Markdown, and plain text. It pulls text from tables in spreadsheets, slide notes in presentations, and metadata from ebook files. If you can store information in it, DocuBrowser can probably index it.

Why This Matters

If you’re a developer, you probably have a folder full of API documentation, technical ebooks, and README files you’ve collected across projects. If you’re an IT manager — which, as someone leading an ICT division, I very much am — that collection balloons to include policy documents, compliance frameworks, vendor specs, and project archives.

The bigger that collection gets, the less useful it becomes. Files you can’t find might as well not exist. DocuBrowser solves this by making every document discoverable through both exact keywords and conceptual meaning. And because everything stays local, there’s no privacy concern — no cloud API processing your internal policy documents. It’s the same privacy-first philosophy behind tools like Meetily for meeting transcription or running your own search infrastructure.

What You’ll Need

• Python 3.9+ — most modern systems have this pre-installed
• Ollama — DocuBrowser installs this automatically if missing, but you can grab it from ollama.com
• A modern browser — Chrome, Firefox, Edge, or Safari
• At least 4GB of free RAM — the embedding model runs locally and needs headroom
• Optional: Calibre — needed for MOBI/AZW3 ebook indexing (sudo apt install calibre or sudo dnf install calibre)

Installation

DocuBrowser ships as ready-to-install packages for Linux (RPM, DEB, tarball), Windows (zip), and macOS (dmg).

Package Install (Recommended)

Download the appropriate package from the GitHub Releases page:

# Debian / Ubuntu / Mint
sudo apt install ./docubrowser-foss_0.9.0-7_all.deb

# Fedora / RHEL
sudo dnf install ./docubrowser-foss-0.9.0-7.noarch.rpm

# Any Linux (tarball)
tar xzf docubrowser-foss-0.9.0-7.tar.gz
cd docubrowser-foss-0.9.0-7
sudo ./install.sh

All three methods install to /opt/docubrowser/ with a Python virtualenv, CLI wrappers at /usr/bin/docubrowser and /usr/bin/docuback, and a desktop menu entry under Office.

Run from Source

If you prefer the repo directly:

git clone https://github.com/linuxrebel/DocuBrowser.git
cd DocuBrowser
pip install -r requirements.txt

Then use ./docubrowser.py instead of the docubrowser command for the rest of this guide.

First Run: Three Commands

Once installed, the entire workflow is three commands:

# Scan and index your documents
docubrowser rescan

# Start the web server
docubrowser start

# Open the UI in your browser
docubrowser open

The first time you run docubrowser start, it automatically checks that Ollama is installed and pulls the two required models — nomic-embed-text:latest for embeddings and dolphin3:latest for AI synopses. This takes a few minutes on the first run. Subsequent startups are near-instant.

By default, it indexes documents in your home directory, but you can point it at any folder through the Settings page. You can even add multiple scan directories that get merged into a single index — perfect if your documents live across a work folder and a personal research folder.

How the Search Works

This is where DocuBrowser stands out. When you type a query, it doesn’t just match words — it understands intent.

• Keyword search: SQLite FTS5 gives you fast, traditional full-text search across titles, authors, subjects, and document snippets. It’s your safety net — exact matches never get lost. If you’ve worked with embedded databases before, the indexing approach has a similar philosophy to tools like LMDB — fast local storage with minimal overhead.
• Semantic search: The nomic-embed-text model converts your query into an embedding vector and finds documents with similar meanings. Search for “that contract about the lease renewal” and find it even if the file is named Commercial_Property_Agreement_2025.pdf.
• Hybrid mode (default): 70% semantic plus 30% keyword weighting, merged and re-ranked. Best of both worlds.

Every result shows a relevance score badge (0–100%), so you can gauge match quality at a glance. Results are paginated at 50 per page.

AI Synopsis: Know Before You Open

Click any document title in the search results, and DocuBrowser generates a Kindle-style book-jacket summary using the dolphin3 model. It tells you what the document is about so you can decide if it’s worth opening — without actually opening it.

This sounds small, but when you’re sifting through a stack of search results, being able to dismiss irrelevant documents in seconds saves a surprising amount of time. The synopsis is cached in SQLite after the first generation, so revisits are instant.

And when you find the right document, the Open button launches it in your default application via xdg-open. No hunting through file managers.

Privacy and Safety Features

For anyone handling sensitive documents — and in government IT, that’s basically all of us — DocuBrowser includes thoughtful safeguards:

• PII scanner: Post-ingest, it scans for Social Security numbers, credit cards, bank accounts, dates of birth, medical records, driver’s licenses, and passport patterns. Matching documents are removed from the index and permanently blacklisted.
• Ignored directories: Exclude specific folders from scanning entirely through the Settings UI.
• Memory safety: Kernel-enforced RLIMIT_AS caps memory at 6GB per worker, with automatic pause and resume on free-RAM thresholds. No OOM crashes on underpowered machines.

These features pair well with other privacy-first tools I’ve covered, like self-hosting SearXNG as your own private search engine — same philosophy of keeping data where it belongs.

Real-World Scenarios

Here’s where I see DocuBrowser fitting into actual workflows:

• Government IT compliance: Index policy documents, DICT memos, and security frameworks. Search by concept — “data privacy breach notification procedure” — even if that exact phrasing doesn’t appear in any single document.
• Graduate research: I’m working on my Master’s thesis right now, and keeping track of 50+ research papers is exactly the problem DocuBrowser solves. Search for “that paper about zero-trust in government networks” instead of scrolling through Dropbox.
• Development documentation: API docs, architecture decision records, onboarding guides — all in one searchable index.
• Personal knowledge management: Markdown notes, saved articles, meeting transcripts. If it’s a document, DocuBrowser can find it.

Honest Trade-Offs

No tool is perfect. Here’s what DocuBrowser doesn’t do well — and you deserve to know before you invest time in it:

• Scanned PDFs (image-only): DocuBrowser detects image-only PDFs and routes them to a separate list. It doesn’t run OCR. If your collection is mostly scanned paper, you’ll need an external OCR pass first.
• DRM-protected ebooks: AZW files with DRM are indexed with metadata only (title, author visible). The body can’t be searched. Fair, but worth flagging.
• Resource usage: Running both nomic-embed-text and dolphin3 locally means you’ll want 8GB of RAM for a comfortable experience. On a 4GB machine, expect slower first-time synopses.
• It’s v0.9.0: Pre-1.0 software. The developer (James Sparenberg) has been shipping consistently — 269 commits, 13 releases — but you should expect occasional rough edges.

Bottom Line

DocuBrowser solves a problem so specific and so universal that I wonder why nobody built it properly before. It’s not trying to replace Google Drive or Notion. It’s filling the gap between “I have too many documents” and “I can actually find what I need.”

Sometimes the best tools aren’t the flashiest. They’re the ones that quietly fix a headache you’d learned to live with.

Image: Code on computer monitor (Unsplash) by Fæ via Wikimedia Commons (CC0)