Local Windows 11 search and launcher in Rust: files, apps, settings and semantic media search with EmbeddingGemma 2.
Python
26
35 commits
updated Oct 15, 2026
A local, universal search and launcher for Windows 11.
VectorDash is meant to be a full Windows search replacement: one search bar for files and folders, installed apps, Windows settings, and quick calculations. Press Ctrl+Shift+Space 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.
| Search for | Example | What happens |
|---|---|---|
| Files and folders | invoice, forzahorizon6.exe | Find names and paths across your selected drives and folders. |
| File contents | a dog on the beach, the clause about pets | Find matching photos, screenshots, document passages, PDF pages, or audio/video moments. |
| Apps | chrome, app: calculator | Launch Start menu programs and Windows apps, with their actual icons in results and previews. |
| Windows settings | bluetooth, settings: sound | Open settings pages such as Wi-Fi, display, storage, sound and Windows Update. |
| Calculations | 24*1.2, 20% of 150, sqrt(81) | Get an instant local answer; press Enter to copy it. |
To enable program launching, open Sources → Start menu programs and turn on Include Start menu programs in search. The catalog refreshes at startup; use Scan Start menu after installing a program. Select a result and press Enter to launch its original shortcut, preserving arguments and working folders.
Names search needs no model. In Sources, choose Quick file / folder names or Full index before adding drives, folders or files. To change existing sources, select entries (or Select all) and apply a mode, then update the index. In either search window, choose Names, Semantic or Both, and use the source dropdown to search some or all sources. More specific folders take priority when sources overlap. Index status shows Full/Quick file counts and how many files have vectors. Enable Skip Windows system and frequently changing cache folders in Sources to omit temp files, crash dumps, update staging and selected caches while keeping personal folders and Program Files.
Semantic search needs the 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.
| Shortcut | Action |
|---|---|
| Ctrl+Shift+Space | Open quick search; change this in Settings |
| Space | Open/close the larger preview |
| Enter | Open a file, launch an app, open settings, or copy a calculator answer |
| Ctrl+C | Copy the selected file, calculator answer or settings link |
| Tab / Shift+Tab | Cycle file categories |
Approximate sizes vary by release. Model sizes are individual files; runtime sizes are compressed downloads. Installed files, the index and RAM/VRAM use require additional space.
| EmbeddingGemma 2 | Text model | Matching image/audio encoder (mmproj) |
|---|---|---|
| Q8_0 — default | 310 MB | 555 MB |
| BF16 | 558 MB | 982 MB |
| Tool | What it does | Approx. size |
|---|---|---|
| llama.cpp / llama-server — CPU | Runs the model using system memory | 20 MB |
| llama-server — Vulkan | GPU/iGPU inference with compatible drivers | 32 MB |
| llama-server — CUDA | NVIDIA GPU inference, including companion runtime | 577–657 MB |
| Tesseract OCR — optional | Reads exact words in screenshots and scanned PDFs | 50 MB |
| OCR language packs — optional | Enables OCR for selected languages | English 4 MB |
| FFmpeg | Decodes audio/video and extracts frames and audio | 88 MB binary; no extra download |
Install only the runtime for your hardware. OCR is optional for image meaning searches; it helps with exact text. FFmpeg and the encoder are needed for media processing.
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 Models & hardware → Download and set up; 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.
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:
| Profile | Search cache limits | Best for |
|---|---|---|
| Low memory | 8 MiB of database pages per search; vectors streamed from disk | Lowest RAM use |
| Balanced — default | 32 MiB of pages per search; up to 128 MiB of compact vectors total | Fast everyday search |
| Fast search | 64 MiB of pages per search; up to 512 MiB of compact vectors total | Larger semantic indexes |
On an NVMe test with 988,658 indexed files and 34,678 vectors, encrypted Balanced searches took 6–14 ms for names and 64–238 ms for semantic retrieval. Main-window/popup results took about 220–407 ms, including typing delay and preview; repeated cached searches displayed in 4–8 ms. Full-app peak RAM was about 336 MiB, excluding the model and processing helpers. Low memory mode used about 230 MiB in the same UI check. These are measured development results, not guarantees. See all profiles, encryption on/off, and indexing measurements.
One- and two-character name searches use filename prefixes; longer queries also match substrings. Broad searches return a bounded selection of best candidates. Cache limits are caps, not reserved memory; models and active file-processing helpers use additional RAM/VRAM.
Database encryption is enabled by default. Change it in Settings → Index privacy, then restart. Encryption protects stored filenames, extracted text and embeddings but can add indexing/search overhead. Disabling it may improve disk performance; anyone with access to that database can read its contents. Existing indexes convert in the background on the first launch after upgrading or changing encryption. A large index can take several minutes to convert once; the window remains usable while file search waits for conversion. No full re-embedding is needed.
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
VectorDash is an independent Windows project inspired by DigUp, created by Abdur Rahim. See acknowledgements. Licensed under MIT.
Local Windows 11 search and launcher in Rust: files, apps, settings and semantic media search with EmbeddingGemma 2.
Python
26
35 commits
updated Oct 15, 2026
A local, universal search and launcher for Windows 11.
VectorDash is meant to be a full Windows search replacement: one search bar for files and folders, installed apps, Windows settings, and quick calculations. Press Ctrl+Shift+Space 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.
| Search for | Example | What happens |
|---|---|---|
| Files and folders | invoice, forzahorizon6.exe | Find names and paths across your selected drives and folders. |
| File contents | a dog on the beach, the clause about pets | Find matching photos, screenshots, document passages, PDF pages, or audio/video moments. |
| Apps | chrome, app: calculator | Launch Start menu programs and Windows apps, with their actual icons in results and previews. |
| Windows settings | bluetooth, settings: sound | Open settings pages such as Wi-Fi, display, storage, sound and Windows Update. |
| Calculations | 24*1.2, 20% of 150, sqrt(81) | Get an instant local answer; press Enter to copy it. |
To enable program launching, open Sources → Start menu programs and turn on Include Start menu programs in search. The catalog refreshes at startup; use Scan Start menu after installing a program. Select a result and press Enter to launch its original shortcut, preserving arguments and working folders.
Names search needs no model. In Sources, choose Quick file / folder names or Full index before adding drives, folders or files. To change existing sources, select entries (or Select all) and apply a mode, then update the index. In either search window, choose Names, Semantic or Both, and use the source dropdown to search some or all sources. More specific folders take priority when sources overlap. Index status shows Full/Quick file counts and how many files have vectors. Enable Skip Windows system and frequently changing cache folders in Sources to omit temp files, crash dumps, update staging and selected caches while keeping personal folders and Program Files.
Semantic search needs the 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.
| Shortcut | Action |
|---|---|
| Ctrl+Shift+Space | Open quick search; change this in Settings |
| Space | Open/close the larger preview |
| Enter | Open a file, launch an app, open settings, or copy a calculator answer |
| Ctrl+C | Copy the selected file, calculator answer or settings link |
| Tab / Shift+Tab | Cycle file categories |
Approximate sizes vary by release. Model sizes are individual files; runtime sizes are compressed downloads. Installed files, the index and RAM/VRAM use require additional space.
| EmbeddingGemma 2 | Text model | Matching image/audio encoder (mmproj) |
|---|---|---|
| Q8_0 — default | 310 MB | 555 MB |
| BF16 | 558 MB | 982 MB |
| Tool | What it does | Approx. size |
|---|---|---|
| llama.cpp / llama-server — CPU | Runs the model using system memory | 20 MB |
| llama-server — Vulkan | GPU/iGPU inference with compatible drivers | 32 MB |
| llama-server — CUDA | NVIDIA GPU inference, including companion runtime | 577–657 MB |
| Tesseract OCR — optional | Reads exact words in screenshots and scanned PDFs | 50 MB |
| OCR language packs — optional | Enables OCR for selected languages | English 4 MB |
| FFmpeg | Decodes audio/video and extracts frames and audio | 88 MB binary; no extra download |
Install only the runtime for your hardware. OCR is optional for image meaning searches; it helps with exact text. FFmpeg and the encoder are needed for media processing.
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 Models & hardware → Download and set up; 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.
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:
| Profile | Search cache limits | Best for |
|---|---|---|
| Low memory | 8 MiB of database pages per search; vectors streamed from disk | Lowest RAM use |
| Balanced — default | 32 MiB of pages per search; up to 128 MiB of compact vectors total | Fast everyday search |
| Fast search | 64 MiB of pages per search; up to 512 MiB of compact vectors total | Larger semantic indexes |
On an NVMe test with 988,658 indexed files and 34,678 vectors, encrypted Balanced searches took 6–14 ms for names and 64–238 ms for semantic retrieval. Main-window/popup results took about 220–407 ms, including typing delay and preview; repeated cached searches displayed in 4–8 ms. Full-app peak RAM was about 336 MiB, excluding the model and processing helpers. Low memory mode used about 230 MiB in the same UI check. These are measured development results, not guarantees. See all profiles, encryption on/off, and indexing measurements.
One- and two-character name searches use filename prefixes; longer queries also match substrings. Broad searches return a bounded selection of best candidates. Cache limits are caps, not reserved memory; models and active file-processing helpers use additional RAM/VRAM.
Database encryption is enabled by default. Change it in Settings → Index privacy, then restart. Encryption protects stored filenames, extracted text and embeddings but can add indexing/search overhead. Disabling it may improve disk performance; anyone with access to that database can read its contents. Existing indexes convert in the background on the first launch after upgrading or changing encryption. A large index can take several minutes to convert once; the window remains usable while file search waits for conversion. No full re-embedding is needed.
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
VectorDash is an independent Windows project inspired by DigUp, created by Abdur Rahim. See acknowledgements. Licensed under MIT.
