VectorDash

A local, universal search and launcher for Windows 11 and Windows 10 (BETA). Built in Rust.

VectorDash is meant to be a full Windows search replacement: one search bar for files/folders, installed apps, Windows settings, and quick calculations. Press Ctrl+Alt+Space (configurable) from any app to find what you need.

Search by name or exact words, or describe what you remember. Google's EmbeddingGemma 2 powers semantic file and media search. Filename search, app launching, settings shortcuts and the calculator work without a model.

What you can search

• Camera RAW (BETA): optional ARW, CR2/CR3, DNG, RAF, NEF and other LibRaw-supported formats use embedded previews first, with optional development fallback; enable in Sources, save and update a Full source, and please test your camera files.
• Interface localization (BETA): 13 languages cover search, settings, help, dialogs and tray menus; native-speaker and complex-script testing is still needed.
• Indexing reliability: file notifications stay targeted, interrupted chunks resume, inaccessible network subfolders do not stop accessible siblings, and completed manual scans satisfy older recovery work.
• Indexing controls: choose Quick or Full first, prepare one batch ahead and share long-file work fairly; optional near-identical frame skipping and learned timing estimates are BETA and need broader media testing.
• Search tuning (BETA): optional FP16 semantic caching reduces vector RAM, while Rank Calibration adjusts cross-media ranking at search time without changing stored embeddings or requiring reindexing; test relevance on your own collection.
• Resizable previews and picture grids: Search and Clipboard dividers remember their widths, Images/Videos grids have a separate preview width, and video thumbnails load on demand into a bounded RAM cache.
• Update channels: choose Release or Beta in About, inspect changes since your installed version, and return to VectorDash automatically after a UI update.
• Remote warm-up and chat models: optionally wake a remote embedding model as Hybrid/Semantic search opens and switch discovered LLM models from the compact dropdown in AI search.
• File search: combine names, exact words and semantic matches. Read PDF, Word, Excel and PowerPoint content; optionally use OCR for screenshot text. Search across 100+ languages and opt into code search.
• Quick search and previews: Hotkey quicksearch from anywhere! Preview results, inspect why they matched, and open, copy or reveal files in Explorer.
• Your sources, your choice: index names only for fast drive-wide search, or fully index selected folders. Nested Full folders override Quick drives. Select multiple search sources, exclude files/types or common Windows cache folders, and see Full/Quick counts and recent additions.
• Advanced and saved searches: combine filenames, wildcards, size and dates, such as name:dog*.jpg size:>1MB or ext:pdf modified:today. Save queries together with their modes, sources and exclusions; rename or remove them in Settings.
• Automatic updates: monitor new, modified and removed files, or schedule scans. Automatic changes obey CPU/GPU, inactivity and power conditions even when periodic scanning is Manual. Optional NTFS journal tracking replays saved changes when volume access is available; folder notifications and recovery scans provide fallback.
• Flexible setup: choose GPU, iGPU or CPU/system memory; install selected optional features or remove downloaded tools. Status shows disk usage and live process RAM/VRAM. Local models warm when Hybrid/Semantic search opens, with loading progress, and can unload after an adjustable idle period. Optional game/full-screen rules pause indexing and unload models; Load anyway overrides model limits for that search session. Use Models and AI for embedding, dictation, LLM and voice models. Remote checks report readiness, vision/mmproj support, output dimensions and available context; manual context limits stay editable. Hardware shows combined local CPU/GPU, memory and storage usage. Media search requires a compatible multimodal embedding server.
• Emoji and clipboard: an offline full-color emoji grid and optional clipboard history for text, links and images, with timestamps, previews and expiry. Optionally search copied images by meaning with EmbeddingGemma 2, just like regular search.

• Voice dictation: optional local Parakeet speech recognition on CPU or NVIDIA GPU, or a remote OpenAI-compatible transcription server. The model loads immediately on activation and unloads immediately or after your chosen idle interval. Set a voice-shortcut popup default (Search, Clipboard or AI), response length and automatic read aloud in Settings.
• Make it yours: light/dark themes, app accent colors, popup tint/transparency/gradient, a collapsible icon sidebar. About includes updates and a changelog covering releases since your installed version.

AI and agent access

Switch to AI to ask questions using your own local or remote OpenAI-compatible server. In Models and AI, enter a base URL such as http://localhost:8080/v1, an optional API key, and your model; refresh to select from the server's models and set a maximum context size.

Enable AI chat and select a model to show the AI tab. Choose Brief (one or two direct, spoken-style sentences - just like a voice assistant), Medium (8–12 sentences) or Full (model default).

With a vision-capable model, attach a monitor, window or selected screen area to ask about what you see.

The optional MCP server, on the same settings page, lets compatible agents search indexed files and retrieve file information. Keep it local or enable access on your LAN, with optional bearer authentication. See MCP setup for connection details.

Install

1. Download the latest Windows x64 setup EXE from Releases (about 106 MB), or the portable ZIP.
2. Run the installer. The EXE is not code-signed, so Windows may show a SmartScreen or unknown-publisher warning. If you trust the release, choose More info → Run anyway, when available, to continue.
3. Choose CPU/GPU settings in Hardware, then open Models and AI → Embedding Model → Download/setup to install the model, encoder and selected runtime. Click Start when needed. Q8_0 is the default. For names-only search, skip model setup and select Quick mode in Sources.
4. Add folders or drives in Sources, then click Update index. Start with a small folder to check results.

Clear an index: use Sources → Clear entire index and confirm. This removes indexed names, extracted text and vectors, while keeping your original files, sources, settings and downloaded models. Automatic indexing stays paused until you resume it or click Update index. Previously retained import/conversion backups are separate and are not deleted.

Enable Include Start menu programs in Sources to search installed program shortcuts. The catalog loads at startup. Select a result and press Enter to launch its original shortcut, preserving arguments and working folders.

File/Folder Names search needs no model. In Sources, choose Quick file / folder names or Full index before adding drives, folders or files. In either search window, choose Hybrid, File/Folder Names or Semantic, and use the source dropdown to search some or all sources.

Semantic search needs the embedding model and a compatible llama-server. Image/audio search also needs its matching encoder. You can select existing files instead of downloading them. Changing the model or encoder requires rebuilding the index once.

Customize hotkeys in Settings → Search and dictation → Keyboard Shortcut.

Windows 10 support is BETA, targeting x64 version 2004/build 19041 or newer; Windows 11 remains the primary tested platform, and actual Windows 10 device testing is still needed.

Models and tools

Approximate sizes vary by release. EmbeddingGemma sizes are individual files; Parakeet and runtime sizes are compressed downloads. Installed files, the index and RAM/VRAM use require additional space.

Parakeet provides English speech-to-text, independently of semantic search. Either model can use CPU or NVIDIA CUDA. Models load on microphone activation; immediate unloading is the default. Installed disk usage is shown in Status.

Install only the runtime for your hardware. OCR is optional for image meaning searches; it helps with exact text. Windows native decoding is the default for video frames, with separate CPU/GPU settings and optional FFmpeg fallback. FFmpeg is still used for audio extraction/playback; semantic media search also needs the matching encoder.

Download trouble? Setup caches official catalogs and retries temporary failures. Rate-limit messages identify GitHub (runtimes) or Hugging Face (models) and suggest when to retry. Optional GitHub / Hugging Face download tokens are in Hardware → Optional download tokens; save them before retrying. They are protected by your Windows account and sent only to their own service. Interrupted downloads retain partial files for resume, and checksums are still required. You can also select an existing runtime/model manually.

Performance

The database and filename catalog stay on disk. Background indexing uses bounded memory and reuses unchanged files. Each source shows its indexed file count and new additions in the past hour.

Choose Settings → Search performance:

On a development NVMe test with 1,076,786 files and 14,123 vectors, encrypted Balanced searches measured about 9 ms for warm filename searches, 32 ms for warm semantic retrieval, and 47 ms for Hybrid. These are backend timings with warm disk caches and an already loaded GPU model; cold model loading, UI delay and hardware add time. An empty native GUI measured about 115 MiB, excluding model/helper memory. See all native profiles, encryption on/off, indexing measurements and limitations.

Database encryption is enabled by default. Settings lets you change it and explains the performance tradeoff. Encryption protects stored filenames, extracted text and embeddings but adds processing overhead. Changing it creates a converted database copy without re-embedding; the original is retained, including any plaintext copy. Portable encrypted data is tied to your Windows account/machine. Existing Rust installations retain their settings/index; older Python users can import a copy through Sources.

Privacy and limitations

Local mode keeps indexed content on your computer; setup downloads need internet access. A remote model server receives the content sent for embedding. Enable LAN/MCP access only for trusted clients. See security details.

AI search can miss or misrank results. OCR can misread text, and sampled video/audio matches give approximate moments. Check original files before relying on a result. VectorDash was developed with AI assistance.

File support · MCP setup · Build from source and migration

VectorDash is an independent Windows project inspired by DigUp, created by Abdur Rahim. See acknowledgements. Licensed under MIT.