VectorDash 0.3.23

Improvements

• Native Windows video frame decoding, with separate Auto/CPU/GPU controls and GPU adapter selection. FFmpeg remains available for compatibility and audio extraction; its worker threads are limited.
• Sources indexing settings grouped into clearly labeled Images, Audio, Video, PDF, Applications/code and shared-limit boxes.
• Faster microphone activation, denser voice waveform and a clear Transcribing status.
• Fixed source-list sizing and left-aligned exclusions with adjacent Remove buttons.
• Polished AI search: model information inside the search bar, themed Send glow, playback controls below answers and correct compact popup height when switching tools.
• Fixed remote multimodal embedding requests and added a real image capability check. Improved stale GPU/backend selection handling.

Validation

130 regression tests passed, plus the native codec test and packaged UI checks. Native video extraction was tested on CPU, NVIDIA RTX 5080 and AMD Radeon integrated graphics without FFmpeg. Decoder benchmarks and their limitations are documented in Retrieval.

Downloads

Use the setup EXE to install or update VectorDash, or the ZIP for portable use. Downloadable models are installed separately on demand. SHA-256 checksums are included.

VectorDash 0.3.22

VectorDash 0.3.22 refreshes the main window and quick search with one seamless frosted-glass surface.

• Larger, smoother rounded corners with the extra square border removed.
• Main and popup windows share color tone, transparency and gradient settings.
• Seamless search field, aligned search icon and full-width divider.
• Tighter spacing and compact popups without empty results or AI-response space.
• Rounded search and clipboard image previews.
• Cycle Search / Clipboard / AI is in the main keyboard shortcuts section.
• New gentle C4 → G4 dictation activation chime.

Validation: 126 automated tests passed; main/popup native window checks passed; the final test build was approved after desktop testing.

Download the setup EXE to install or update, or the ZIP for portable use. Windows 11 x64. The installer is unsigned, so Windows may display a warning. Existing settings and index are preserved.

Experimental RAW-photo support remains in its separate preview release and is not included here.

VectorDash 0.3.21 — RAW preview 1

Test build for PR #13, starting camera RAW support requested in issue #12. This is a prerelease; 0.3.20 remains the official stable release.

Try it

1. Install the RAW Preview Setup below, or extract the portable ZIP.
2. In Sources, choose a Full photo source and enable Index image meaning and Index camera RAW photos. Save, then Update index. RAW support defaults off, including on upgrades.
3. For files without an embedded preview, separately enable Develop RAW when no usable preview exists (also off by default).
4. Search Images, use Space for Quick Look, and Open/Reveal/Copy the original RAW. Quick sources stay filename-only. Semantic indexing needs the usual multimodal embedding server/encoder.

Included

• Embedded JPEG/RGB previews first; optional reduced-resolution LibRaw development.
• ARW, CR2/CR3, DNG, RAF and other supported camera formats; original files remain untouched.
• RAW-specific upgrade stamps, corrupt-file warnings/name-search fallback, Unicode paths and orientation handling.
• Separate RAW size limit (256 MiB default), single decoder queue, 512 MiB worker cap, 45 seconds per attempt and 2048px output. Development is limited to 50MP sensors. Decoder exits after use.

Validation and limits

131 automated tests passed. Packaged native indexing was checked with real CC0 Sony, Canon, Fujifilm and Blackmagic fixtures, including a RAW over 32 MiB and a DNG without a usable preview. Unchanged rescans made no embedding calls; Quick sources never decoded RAW; original hashes were unchanged. Worker cancellation left no decoder process. Embedding plumbing used a deterministic local fixture server, so these are not visual-relevance accuracy benchmarks.

On one 34 MB ARW, direct preview extraction was about 39 ms / 8 MiB decoder RAM, compared with about 226 ms / 125 MiB for development; excludes model inference and cold I/O. The small packaged indexing fixture peaked around 51 MiB across app/extraction helpers, excluding its external embedding server. Camera variants and files exceeding resource limits may remain filename-only.

The PR contains the implementation plan. A bounded cache that follows index encryption and broader camera/relevance testing remain planned; this build writes no persistent derived-photo cache. LibRaw source and notices are included.

The installer is unsigned; Windows may display a SmartScreen/unknown-publisher warning. Preview asset names intentionally differ from stable assets so older stable auto-updaters will not offer this experimental build.

VectorDash 0.3.20

Official Windows release.

• Improved Quick Look geometry, query selection, compact clipboard results and literal-text matching.
• AI response styles, follow-up chat, screenshot attachments and per-response copy/read-aloud controls.
• Optional TTS with voice discovery, connection tests, streaming where supported and automatic playback controls.
• Voice-shortcut defaults; local or remote embeddings and dictation; model discovery and capability checks.
• Hardware resource status, separate model/runtime uninstall controls, and reorganized Models and AI settings.
• Configurable Ctrl+Tab file-type cycling and grouped keyboard shortcuts.

Validation: GitHub native build and regression checks passed; 125 automated tests plus isolated main/popup and playback fixtures.

Download Setup to install/update, or ZIP for portable use. The installer is unsigned, so Windows may show a SmartScreen warning. Models download separately.

VectorDash 0.3.14 Prerelease

VectorDash 0.3.14 prerelease

• Full-width microphone waveform inside the search input, in the main window and popup. Recording instructions are hidden; useful transcription errors remain visible.
• Transcription removes sentence-ending punctuation while preserving punctuation inside your query. AI dictation still submits automatically.
• Recording trail is 50% longer, with red fading into orange over its last half. AI processing uses the same trail size with a purple-to-blue gradient.
• New Enable AI chat setting. AI stays hidden when disabled or no model is selected; Clipboard stays hidden when history is disabled. Tab and Shift+Tab skip hidden modes. MCP remains independent.
• Popup footer shows search feedback instead of background indexing summaries.

Validation: 101 automated tests passed, plus native main/popup checks for waveform rendering, both trails, punctuation cleanup, AI auto-submit, hidden modes and popup status filtering. AI tests used a local fixture server.

Windows 11 x64. Setup and portable downloads include SHA-256 checksums. Setup is unsigned and may trigger a Windows warning. This is a prerelease; v0.3.12 remains the latest stable release.

VectorDash 0.3.13 Prerelease

VectorDash 0.3.13 prerelease

• Full-color, offline emoji in the selector and AI replies. Copying keeps the original Unicode emoji. Bundled Twemoji artwork is credited and licensed in the installer and portable ZIP.
• Native rounded acrylic popup corners, replacing the rectangular legacy blur surface. Popup painting follows the same native corner radius; the main window retains its existing rounded shape.
• Readable amber Installed · on demand text in light mode, improved status/warning colors, and red borders for Uninstall all downloaded models and runtimes and Clear entire index.
• Dictation shortcuts preserve the active Search, Clipboard, AI or Emoji mode when a search window is open. AI dictation submits automatically after transcription.
• Visible windows paint immediately, including when initially opened without keyboard focus.

Validation: 96 automated tests passed. Native UI tests confirmed color emoji rendering, mode retention and AI auto-submit against a local fixture server. Popup corner rendering was compared over a plain white test window.

Windows 11 x64. Setup and portable downloads are included with SHA-256 checksums. The setup EXE remains unsigned, so Windows may show a SmartScreen warning. This is a prerelease for testing; v0.3.12 remains the latest stable release.

VectorDash 0.3.12

VectorDash 0.3.12 adds voice dictation, clipboard history and AI tools while fixing excessive CPU use in the system tray.

New features

• Optional local English dictation with NVIDIA Parakeet TDT 0.6B v2: CPU/CUDA, fp16/8-bit, microphone selection, live waveform and automatic speech-end detection. Loads on microphone activation; unload immediately or after an idle interval.
• Soft microphone activation sound and a smooth glowing red recording border.
• Offline emoji grid and optional protected clipboard history for text, links and images, with timestamps, previews and expiry. Optional semantic search for copied images.
• OpenAI-compatible AI chat with model selection, Markdown, follow-ups, cancellation and Ctrl+C answer copying. Vision models can receive monitor/window screenshots or directly selected screen areas.
• Remembered window positions, improved filters and saved-search layout, and clearer model status with RAM/VRAM, disk usage and install/uninstall controls.

Fixes and improvements

• Fixed continuous repainting in tray mode, cutting measured idle CPU use by 99.8% in a development test.
• Preserved existing global shortcuts across updates; added registration recovery. New installations default to Ctrl+Alt+Space for quick search and Ctrl+Shift+D for dictation.
• Tab / Shift+Tab cycle Search, Clipboard and AI. Renamed Both to Hybrid.
• Improved targeted file-change indexing, pause/cancel handling and full-screen detection.
• Fixed managed model removal, including legacy downloads and Windows short-name user paths.

Downloads

Use the setup EXE to install or upgrade, or the portable ZIP. Optional models download separately. Windows 11 x64 is required. The installer is unsigned, so Windows may show a SmartScreen warning.

Validation: 93 automated tests passed, plus real CPU/CUDA dictation, process reuse and unload checks. SHA-256 checksums are included below. See the updated README for model sizes and usage.

VectorDash 0.3.11

VectorDash 0.3.11 adds search model controls, a full index reset and safer scan cancellation.

• Both/Semantic search warms local EmbeddingGemma 2 and llama-server, with loading progress in main and popup search. File/Folder Names needs no model.
• Optional game/full-screen rules pause indexing and unload local models. Optional CPU/GPU/power limits defer search loading. Load anyway protects the model for that search session; closing search releases it.
• Automatic game detection uses Windows signals and common Steam/Epic/GOG game paths. Add executable names for games those checks miss.
• Sources: Clear entire index removes indexed names, extracted text and vectors with confirmation. Your files, sources, settings and downloaded models are kept. Automatic indexing remains paused until resumed; retained import/conversion backups are separate.
• Canceling a scan no longer immediately starts recovery. Saved source changes stop and restart indexing with the current sources.

72 tests passed, including encrypted/plaintext index reset, model readiness and unload protection, automatic change indexing and scan cancellation. Setup and portable binaries/checksums verified.

Close VectorDash before installing. Download the setup EXE or portable ZIP below. Setup is unsigned; Windows may show a SmartScreen/unknown-publisher warning.

User guide | Release details

VectorDash 0.3.9

VectorDash 0.3.9 fixes Windows branding and installation upgrades.

• Installs as VectorDash, using the original installer identity and directory to upgrade the older Python installation.
• Removes a separately registered Rust preview after successful installation.
• Preserves settings, downloaded models and indexes; older Python indexes remain available for copy import through Sources.
• Windows executable product/file descriptions now display VectorDash.
• Preserves preview startup behavior and removes obsolete startup entries when changed.

65 native regression tests passed. Isolated installer upgrades verified app identity, original install location, preview removal, Python-runtime cleanup and data preservation. Setup/portable binaries and SHA-256 checksums were verified.

Close VectorDash before installing. Download the setup EXE or portable ZIP below. The setup remains unsigned; Windows may display a SmartScreen/unknown-publisher warning. The Rust label in asset filenames is retained for compatibility with existing updaters; the installed program is VectorDash.

User guide · Release details

VectorDash 0.3.8

VectorDash 0.3.8 is an official release.

• Advanced filename/wildcard, extension, size and modified-date filters.
• Saved searches and multi-source selection; controls aligned beside the filters.
• Optional NTFS change-journal replay with folder-notification/recovery fallback. Automatic changes respect idle limits, including with Manual periodic scheduling.
• Fixed main and popup window dragging, Tab focus, dark sliders and popup clipping.
• Improved matched-moment audio playback and visible playback errors.
• Clearer monitoring/idle-wait logs and reliable cleanup of hidden test instances.
• Restored the detailed README with current features, install instructions, keyboard shortcuts, model/tool sizes and measured performance.

Validation: 65 native regression tests passed locally. Real Windows file-notification fallback, new/modified indexing, main/popup drag tests, audio playback and clean package checks passed.

Download the setup EXE or portable ZIP below. Close VectorDash before installing. The installer is unsigned, so Windows may display a SmartScreen/unknown-publisher warning. Existing Rust settings/index are retained; older Python indexes can be imported as a copy. Models are separate downloads.

NTFS journal tracking requires an existing active journal and volume-read permission. Unsupported/inaccessible journals use notifications and recovery scans; a privileged journal service is not included. Initial/recovery scans may still visit all sources. Full privileged live journal replay remains unverified.

Updated user guide · Journal setup and release details · Security