SilkStack Image Browser v2.2.0

The v2.2.0 release introduces AI Search — fully local semantic search and LLM-powered auto-tagging. Every AI feature runs entirely on your machine via WebGPU: your images, prompts, and tags never leave your computer. Both features are premium-gated, and the AI engine (ai-intelligence) powers them with a shared, on-device WebLLM runtime.

Semantic Search

Search your library by meaning, not just keywords. A local Qwen3-Embedding model (0.6B / 4B / 8B — you pick in Settings) turns every image and its tags into a semantic index, so a query like "red fox in snow" finds conceptually related images even when the words never match.

• Semantic mode in the search bar — a one-click toggle switches the search box between keyword and semantic matching, with live status (loading / ready / error).
• Union with keyword search — semantic results appear first, keyword-only matches are appended, and duplicates are removed, so you never lose access to your old search behavior.
• Sparkle badge — semantic hits are marked with a ✨ badge in both the image grid and the table view, so you can see why a result matched.
• Relevance sort — a new sort order ranks results by semantic relevance while a semantic search is active.
• Search-quality tuning — model-aware match thresholds plus a lexical blend with exact-word overlap and synonym/hypernym expansion ("animal" → dog, cat, fox…) make results useful on real corpora.
• One-time indexing — each image is embedded exactly once, results persist in the local database (IndexedDB), and indexing resumes where it left off across app restarts. Vectors load from disk, not re-computed.
• Model choice — Qwen3-Embedding 0.6B / 4B / 8B, with snowflake-arctic-embed fallbacks; the 4B model is the default.

Auto-Tagging with Local LLMs

Auto-tagging now uses a real language model running on your GPU — not just keyword heuristics:

• Qwen3-powered tag extraction — the LLM reads the embedded prompt of each image and extracts descriptive tags: subjects, styles, lighting, concepts. Models from Qwen3 0.6B up to 8B (Llama 3.2 1B/3B and Hermes 3 3B/8B remain available).
• One call, one bounded list — each image is tagged in a single model call producing one flat list (max 15 tags) that includes search-friendly synonyms and alternate phrasings, so the tags match what a searcher would actually type.
• Processed-once guarantee — images carry an isAutoTagged flag; re-runs only process new images, and empty results are not retried endlessly. Clear Auto-Tags resets the flag so you can re-evaluate the whole library.
• Rule-based fallback — if your hardware doesn't support WebGPU or a model fails to load, the system falls back to the rule-based extractor (unless disabled in Settings).
• Auto-tagging and semantic search share a single WebLLM engine — one model loaded at a time per purpose, with a single worker per source.

AI Model Management

• Footer AI-model pill — the footer shows each loaded AI model with its declared VRAM footprint, with load/eject controls. One eject unloads a model; the next use reloads and restores it automatically.
• Settings → AI Intelligence — pick your embedding model and auto-tagging model, toggle semantic search, and watch indexing progress and errors live.

GPU Selection in Settings

• Detected GPU names — the GPU dropdown lists the actual adapters on your machine, so you can pick the discrete GPU on dual-GPU laptops (the WebGPU API can't select an adapter by name — this maps your choice to the right powerPreference).
• Preference options — auto / high-performance / low-power / software.
• Higher buffer limits — a requestDevice shim raises WebGPU buffer limits to the adapter's caps, so larger models (e.g. Qwen3-Embedding 8B) fit where the default 1 GiB cap would reject them.
• Fixes DXGI_ERROR_DEVICE_HUNG-style failures caused by WebLLM silently running on the integrated GPU — set Settings → GPU → High performance if you hit GPU crashes.

Reprocess Images

• Settings → Reprocess Images — wipe all derived data (thumbnails, auto-tags, semantic vectors) and rebuild the library from scratch through the full pipeline: re-scan, re-tag, re-index. Disabled while any processing is in flight.

Premium Gating

• The whole AI Intelligence section (semantic search + auto-tagging) is gated behind a premium license. Purchase a premium license (lifetime license) to use these features.

Developer Tools

• DevTools shell — a new developer window with a Semantic Search Tester and an Auto-Tagging Tester: load models manually, run search queries, inspect similarity scores, and exercise the tagging pipeline against an isolated test store (your library database is never touched in tester mode).

Under the Hood

• IndexedDB v8 migration — a new semanticVectors store; the app migrates automatically on first launch after upgrade, and the empty store costs nothing until semantic search is enabled.
• Processing queue & sequencer — derived-data pipelines run through a bounded queue with a sequencer; tagging and semantic indexing are guaranteed to run once per image even across retries.
• Worker-based AI runtime — the ai-intelligence engine is consumed from its built bundle and runs in a dedicated worker; model loads skip when the app starts offline.
• Large test-suite expansion — semantic store/UI/engine tests, premium-gating and license tests, GPU classification tests, reprocess and processing-queue tests, and more (≈9.5k new lines across both repos).

Bug Fixes & Polish

• Semantic search no longer activates by itself — the search mode now respects the settings toggle and stays off until you enable it.
• Clear semantic search now fully resets the index state consistently (search box, badges, and persisted state agree afterwards).
• Relevance sort fixes — the sort option appears only while semantic hits are on screen and restores the durable sort when cleared.
• Offline launches no longer attempt AI model loads or error out at startup without a connection.
• Semantic search tester fixes — thumbnail previews reuse the app's thumbnail cache, and persisted semantic IDs are matched against both historical and current id shapes.

Feedback

Found a bug or have a feature request? Open an issue!

---