desksearch 0.6.3

Private semantic search engine for your local files, powered by Starbucks 2D Matryoshka embeddings

DeskSearch

Search your files by meaning. Faster than Spotlight. 100% private.

0.83ms search · 13MB binary · 100% offline · 50k+ files · 10× ONNX speedup

---

DeskSearch is a local semantic search engine that understands what you're looking for — not just the words you type. Index your documents, code, and emails in seconds. Search in under a millisecond. Everything stays on your machine.

Available as a Python package (pip install desksearch) and a standalone Rust binary (13MB, zero dependencies).

More screenshots

Settings — Speed Tiers & File Types

Data Sources — Connectors

File Explorer

---

Features

	Feature	What it means
🔍	Semantic search	Finds "Q3 earnings" when you search "quarterly revenue" — understands meaning, not just keywords
⚡	Sub-millisecond search	0.83ms p50 latency powered by Starbucks 2D Matryoshka embeddings
🔒	100% private	All processing runs locally. No cloud. No telemetry. No data ever leaves your machine
📄	30+ file types	PDF, DOCX, PPTX, XLSX, code (20+ languages), email, Jupyter notebooks, archives
👁	Live reindexing	Built-in file watcher auto-detects changes — your index is always current
🎨	Beautiful web UI	React frontend with dark mode, live search, file preview, and keyboard shortcuts
🦀	Rust core	13MB self-contained binary with PDF/DOCX/PPTX/XLSX parsers, hybrid BM25+dense search, embedded frontend
☕	3 speed tiers	Fast (2-layer, 32d) · Middle (4-layer, 64d) · Pro (6-layer, 128d) — you pick the trade-off
🔌	Connector plugins	Local files, email (.eml/.mbox), Chrome bookmarks, Slack exports — extensible architecture
⚙️	ONNX acceleration	10× embedding speedup (171 chunks/sec) with INT8 quantization support
📊	Advanced filters	Filter by file type, date range, size · Sort by relevance, date, size, name · Export as JSON/CSV/text
⭐	Favorites & recents	Bookmark important files and track recently opened documents

---

Quick Start

Python

pip install desksearch
desksearch
# → opens http://localhost:3777

Rust

curl -fsSL https://github.com/wshuai190/desksearch/releases/latest/download/desksearch -o desksearch
chmod +x desksearch
./desksearch
# → opens http://localhost:3777

On first run, DeskSearch walks you through an onboarding wizard to pick folders and a speed tier. After that, one command is all you need.

---

---

CLI Examples

# Semantic search from the terminal
desksearch search "machine learning papers"
desksearch search "budget spreadsheet" --type xlsx --json

# Index specific folders
desksearch index ~/Projects ~/Research

# Check index health
desksearch status
desksearch doctor            # full health check

# Manage watched folders
desksearch folders add ~/Notes
desksearch folders list

# Switch speed tier
desksearch config set search_speed pro

# Run as a background daemon
desksearch daemon start
desksearch daemon install    # auto-start on login (macOS LaunchAgent)

# Benchmark your setup
desksearch benchmark --files 1000

All commands support --json for scripting and automation.

---

DeskSearch vs. Alternatives

	DeskSearch	Spotlight	Everything	Alfred
Search type	Semantic + keyword	Keyword	Filename only	Keyword
Understands meaning	✅ Yes	❌	❌	❌
File content search	✅ 30+ formats	Limited	❌	Via plugins
Search latency	~1ms	~50ms	~1ms	~50ms
Privacy	100% local	Local (Siri opt-in)	Local	Local
Code-aware	✅ 20+ languages	Minimal	❌	❌
Extensible	Plugins + REST API	❌	❌	Workflows
Open source	✅ MIT	❌	❌	❌
Cross-platform	macOS, Linux	macOS only	Windows only	macOS only

---

Powered by Starbucks Embeddings

DeskSearch uses the Starbucks 2D Matryoshka embedding model, which enables flexible layer × dimension truncation for speed/quality trade-offs.

Paper: Starbucks: Improved Training for 2D Matryoshka Embeddings Shengyao Zhuang*, Shuai Wang*, Fabio Zheng, Bevan Koopman, Guido Zuccon — ECIR 2026 (*equal contribution)

Unlike traditional embeddings that require fixed dimensions, Starbucks lets you choose both the number of transformer layers AND the embedding dimension at inference time — no retraining needed. DeskSearch leverages this to offer three speed tiers from a single model.

If you use DeskSearch or the Starbucks model in your research, please cite:

@inproceedings{wang2026starbucks,
  title={Starbucks: Improved Training for 2D Matryoshka Embeddings},
  author={Zhuang, Shengyao and Wang, Shuai and Zheng, Fabio and Koopman, Bevan and Zuccon, Guido},
  booktitle={ECIR},
  year={2026}
}

---

Architecture

DeskSearch runs hybrid retrieval — every query hits a Tantivy BM25 index and a FAISS dense vector index in parallel, then merges results via Reciprocal Rank Fusion (RRF). Embeddings come from the Starbucks 2D Matryoshka model with layer and dimension truncation, running on ONNX Runtime for 10× speedup over PyTorch (171 chunks/sec vs 17). The indexing pipeline parses files across 6 parallel workers, chunks at sentence boundaries, and embeds in batches of 256 for maximum throughput. A connector plugin system (v0.6.0) lets you pull in data from local files, email, Chrome bookmarks, and Slack exports via a unified API.

Your Files (PDF, DOCX, Markdown, Code, ...)
              │
              ▼  Parse → Chunk → Embed
   ┌──────────────────────────────────────┐
   │   30+ parsers → 512-char chunks      │
   │   → Starbucks 2D Matryoshka (ONNX)   │
   └───────────┬──────────────────────────┘
               │
       ┌───────┴────────┐
       ▼                ▼
  BM25 (Tantivy)   FAISS (dense)
  keyword index     semantic index
       │                │
       └───────┬────────┘
               ▼
    Reciprocal Rank Fusion
               │
               ▼
     Ranked Results + Snippets

---

Connectors (v0.6.0)

DeskSearch supports pluggable data connectors to index content from multiple sources:

Connector	What it does
Local files	File system scanning with scheduled sync and live re-indexing
Email	Parse .eml and .mbox files with sender, subject, and date extraction
Chrome bookmarks	Read your Chrome profile's bookmark hierarchy
Slack export	Import Slack ZIP exports with username resolution

Manage connectors via the API (/api/connectors/v2/) or the web UI settings panel. The ConnectorRegistry handles discovery, configuration, and sync scheduling.

---

Speed Tiers

Tier	Layers	Dimensions	Best for
fast	2	32	Large corpora, older hardware
middle	4	64	Default — balanced speed and quality
pro	6	128	Best accuracy, research use

desksearch config set search_speed pro

---

API

DeskSearch exposes a REST API on localhost:3777:

curl "http://localhost:3777/api/search?q=quarterly+revenue&limit=5"
curl "http://localhost:3777/api/status"
curl "http://localhost:3777/api/health"

Python SDK

from desksearch import DeskSearch

with DeskSearch() as ds:
    results = ds.search("quarterly revenue", limit=5)
    for r in results:
        print(f"{r.rank}. {r.filename} ({r.score:.3f})")
        print(f"   {r.snippet}\n")

---

Contributing

Contributions welcome. DeskSearch is MIT-licensed.

git clone https://github.com/wshuai190/desksearch.git
cd desksearch
pip install -e ".[dev]"
pytest

---

License

MIT © Shuai Wang