Vector Search

Semantic similarity search for Obsidian vaults using only on-device vector embeddings. Find related notes automatically as you browse, or run ad-hoc hybrid queries from the sidebar. Everything runs locally with no API keys or cloud services required.

Uses Orama for hybrid full-text + vector search and Transformers.js for on-device embeddings.

Install

Note: This plugin is not yet available in the Obsidian Community Plugins directory (pending review). Use BRAT or a manual install for now.

BRAT (recommended)

1. Install BRAT from Community Plugins
2. Open Settings > BRAT > Add Beta Plugin
3. Enter sdiehl/obsidian-vector-search
4. Enable "Vector Sidebar" in Settings > Community Plugins

BRAT will keep the plugin updated automatically when new releases are published.

Download Release

Download vector-sidebar.zip from the latest release, unzip it into your vault's .obsidian/plugins/ folder, and enable "Vector Sidebar" in Settings > Community Plugins.

From Source

git clone https://github.com/sdiehl/obsidian-vector-search.git
cd obsidian-vector-search
npm install && npm run build
OBSIDIAN_VAULT=/path/to/vault npm run deploy

Re-run npm run deploy after pulling updates. The deploy script also adds the index file to your vault's .gitignore.

Usage

Open the sidebar via the ribbon icon or Cmd+P > "Open vector search sidebar".

• Similar notes: Navigate to any note to see semantically related notes ranked by similarity.
• Hybrid search: Type a query in the search bar and press Enter. Combines keyword matching with semantic similarity for accurate results. The balance between keyword and semantic scoring is configurable.
• Auto-indexing: Notes are re-indexed on save. New and renamed notes are indexed automatically.
• Rebuild: Use Settings > Vector Search > Rebuild to re-index the entire vault.

Indexing Modes

Mode	Behavior
On save	Re-embeds notes when you navigate away after editing
Interval	Periodic full re-index at a configurable interval
Manual	Only indexes when you click "Rebuild" in settings
Read-only	Uses a pre-built index, no writes (for iPad/mobile)

iPad / Mobile Setup

1. Build the index on desktop (or use the CLI: node scripts/index.mjs --vault /path/to/vault)
2. Sync the index file via Obsidian Sync or iCloud
3. On iPad, set indexing mode to Read-only and enable Low memory mode

Low memory mode skips caching vectors in RAM (Orama handles storage internally), roughly halving memory usage. The similar-notes view will embed the active note on each switch instead of using a cached vector.

Embeddings

Embeddings are computed via Transformers.js running ONNX models in a sandboxed iframe using the WebAssembly backend (no native dependencies, works on desktop and iPad).

Model	Dimensions	Quantization	Download	RAM
all-MiniLM-L6-v2 (default)	384	INT8	~23 MB	~50 MB
all-MiniLM-L12-v2	384	INT8	~33 MB	~70 MB
bge-small-en-v1.5	384	INT8	~33 MB	~70 MB

Model files are downloaded from HuggingFace on first use and cached in IndexedDB. WebGPU acceleration is used when available. Falls back to WASM single-threaded on iPad and environments without SharedArrayBuffer.

CLI Indexing

Pre-index a vault from the command line (useful for large vaults or iPad sync):

node scripts/index.mjs --vault /path/to/vault

Releasing

node scripts/release.mjs 0.3.0

License

MIT