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).

Features

Quick Start

Python

pip install desksearch
desksearch
# → opens http://localhost:3777 with onboarding wizard

On first run, DeskSearch auto-downloads the Starbucks embedding model and exports an optimised ONNX (INT8) version. This takes ~30 seconds once; subsequent starts are instant.

For full Starbucks model support (auto-export), also install torch:

pip install desksearch[full]   # includes torch + onnx for auto-export

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

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:

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

Speed Tiers

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