desksearch 0.1.4

Private semantic search engine for your local files

DeskSearch

Search your files by meaning, not just keywords. 100% local, zero cloud.

---

What is DeskSearch?

A private search engine for your local files. It combines keyword matching (BM25) with semantic search (dense vectors) so you can find files by what they mean, not just what they say — and nothing ever leaves your machine.

Install

Desktop App (recommended)

Download the latest release for your platform:

• macOS — .dmg (Apple Silicon)
• Windows — .exe installer
• Linux — .AppImage

👉 Download from Releases

Just open the app — no Python or terminal needed. It handles everything automatically.

pip (for CLI / developer use)

pip install desksearch
desksearch

On first run, the setup wizard detects your document folders, indexes them, and opens a web UI at localhost:3777.

Features

• Desktop app — standalone app with system tray, global hotkey (Cmd+Shift+Space / Ctrl+Shift+Space), and built-in folder browser
• Hybrid search — BM25 (via tantivy) + dense retrieval (FAISS), merged with Reciprocal Rank Fusion
• Fully private — everything runs locally. No cloud, no API keys, no data leaves your machine
• Fast — keyword search in <10ms, semantic search in <200ms
• 30+ file formats — PDF, DOCX, Markdown, HTML, Jupyter notebooks, source code, CSV, LaTeX, and more
• Web UI — clean dark-mode interface with folder browser, file explorer, and real-time indexing
• CLI — search, index, and manage everything from the terminal
• Background daemon — watches folders for changes and re-indexes automatically
• Plugins — extend with custom parsers, rerankers, or data connectors

How It Works

  Your Files (PDF, DOCX, Markdown, Code, ...)
                    │
                    ▼
        ┌───────────────────────┐
        │   Parse → Chunk → Embed  │
        └─────┬───────────┬─────┘
              │           │
              ▼           ▼
        ┌─────────┐ ┌──────────┐
        │  BM25   │ │  Dense   │
        │(tantivy)│ │ (FAISS)  │
        └────┬────┘ └────┬─────┘
             │           │
             └─────┬─────┘
                   ▼
          Reciprocal Rank Fusion
                   │
                   ▼
          Ranked Results + Snippets

1. Parse — extract text from 30+ formats
2. Chunk — split into overlapping passages (512 chars, 64 overlap)
3. Embed — generate 384-dim vectors with all-MiniLM-L6-v2 (runs locally)
4. Search — query both indexes in parallel, fuse results with RRF

CLI Usage

desksearch                          # Start web UI (setup wizard on first run)
desksearch search "your query"      # Search from terminal
desksearch search "ML papers" -n 5  # Limit results
desksearch index ~/Documents        # Index a folder
desksearch status                   # Show index stats
desksearch daemon start             # Run in background with file watcher
desksearch config                   # View/edit configuration

Configuration

Config lives at ~/.desksearch/config.json:

Setting	Default	Description
index_paths	~/Documents, ~/Desktop	Folders to index
embedding_model	all-MiniLM-L6-v2	Embedding model
chunk_size	512	Characters per chunk
port	3777	Web UI port
max_file_size_mb	50	Skip files larger than this

Plugins

Three extension points: parsers (new file formats), search (rerankers), and connectors (external data sources).

Drop a .py file into ~/.desksearch/plugins/:

from desksearch.plugins.base import BaseParserPlugin
from pathlib import Path

class EpubParser(BaseParserPlugin):
    name = "epub-parser"
    extensions = [".epub"]

    def parse(self, file_path: Path) -> str:
        ...
        return text

Performance

Tested on MacBook Pro M1, 10,000 files (~2GB):

Metric	Value
BM25 search	<10ms
Semantic search	<200ms
Indexing	~500 files/min
Memory (idle)	~80MB
Cold start	~1.5s

Building from Source

Desktop App

Requires Python 3.10+ and Node.js 18+.

# 1. Install Python dependencies
pip install -e ".[dev]"

# 2. Bundle the backend
pip install pyinstaller
pyinstaller desksearch.spec  # or see scripts/build-app.sh

# 3. Build the Electron app
cd electron && npm install && npm run dist:mac  # or dist:win / dist:linux

Development

pip install -e ".[dev]"
pytest

# Frontend dev (hot reload)
cd src/ui && npm install && npm run dev

License

MIT

---

Built by Shuai Wang