Vault Search

Local-first semantic search & discovery for Obsidian — simple, private, Chinese-friendly

Vault Search helps you search by meaning and rediscover forgotten notes.

No cloud services. No API keys. No subscription fees. Your notes never leave your machine.

Why Vault Search?

Andrej Karpathy shared his vision of LLM-maintained knowledge bases — letting AI "compile" your notes into a structured wiki. It's a compelling approach, but it assumes you're ready to hand full editorial control to an LLM.

Vault Search takes a different stance. AI should help you see, not think for you. The best tool doesn't replace your writing — it helps you rediscover what you already know, and surfaces connections you missed.

What sets Vault Search apart

Discover, not organize — Other tools build AI wikis or auto-summaries. Vault Search finds the notes you should be looking at. The Discover tab shows related notes you haven't connected yet — especially Cold (isolated) notes hiding in your vault.

Hot/Cold intelligence — Notes with links or recent activity are Hot. Orphan notes are Cold. Discover surfaces Cold notes that are semantically related to your current thinking — your blind spots become visible.

MOC generation — One click turns search or discover results into a Map of Content note with wikilinks and previews. You decide the structure; AI just gathers the pieces.

Truly local, truly private — All embedding, indexing, search, and discovery happen on your machine. Zero data leaves your computer. This isn't a toggle; it's the architecture.

Simple and fast — Sidebar panel with Search and Discover tabs. Cmd/Ctrl+P for instant modal search. Right-click results for Obsidian's native file menu. Drag results to Canvas for visual mapping.

Optimized for Chinese — Built with qwen3-embedding:0.6b, which excels at Traditional Chinese + English semantic understanding. Combined with synonym expansion, even different phrasings of the same concept will match.

LLM-powered descriptions — A local LLM generates frontmatter descriptions for your notes, giving the embedding model a high-quality summary to work with. This dramatically improves search and Discover relevance for long notes.

Runs on 8GB laptops — Minimal memory and CPU footprint. Recommended models work on a MacBook M2 with 8GB RAM. Incremental indexing + debounce means near-zero overhead in daily use.

Flexible and compatible — Works with Ollama, LM Studio, llama.cpp, vLLM, or any OpenAI-compatible server. Choose the models that work best for your language and hardware.

"AI helps you see. You decide what it means."

Features

Search

Semantic Search — Find notes by meaning, not just keywords
Sidebar Panel — Persistent results with Search and Discover tabs
Quick Modal — Cmd/Ctrl+P for fast note jumping
Find Similar — Discover related notes instantly (zero API calls)
Smart Indexing — Incremental updates, auto-indexes on file changes
Hot/Cold Tiers — Hot = linked/recent, Cold = isolated/orphan
Chunking — Optional overlapping chunks for long documents

Discover (v0.3.0)

Active Discovery — Open a note, sidebar auto-shows related notes with Cold notes highlighted
Global Discover — Find Cold notes most related to your Hot (active) notes
MOC Generation — Export search or discover results as a Map of Content note
Cold Search Scope — Dedicated "Cold only" search mode for intentional exploration
Tier Badges — Visual markers distinguish Hot and Cold results at a glance
Canvas Integration — Drag any result directly onto Canvas for visual mapping
Context Menu — Right-click results for Obsidian's native file menu (Bookmark, etc.)

Description Generator

LLM Descriptions — Local LLM generates frontmatter descriptions
Synonym Expansion — Define synonyms to improve recall
Bilingual UI — English & Traditional Chinese (auto-detected)

Requirements

Ollama installed and running
An embedding model (e.g., ollama pull qwen3-embedding:0.6b)
An LLM model for description generation (e.g., ollama pull qwen3:1.7b) (optional)
Obsidian desktop

Installation

BRAT (recommended while pending community review)

Install BRAT plugin
Add this repository: notoriouslab/vault-search
Enable "Vault Search" in Community plugins

Manual

Download main.js, manifest.json, styles.css from the latest release
Copy to .obsidian/plugins/vault-search/ in your vault
Enable in Settings → Community plugins

Note: If your vault is tracked by Git, add .obsidian/plugins/*/data.json to .gitignore to avoid accidentally committing API keys or personal settings.

Quick Start

Settings → Vault Search → Select your embedding model
Click Rebuild to index your vault
Cmd/Ctrl+P → "Semantic search" or click the compass icon
Switch to the Discover tab to see related notes for the current file

Recommended Workflow

1. Generate descriptions  →  2. Rebuild index  →  3. Search & Discover
   (LLM summarizes notes)    (embed with descriptions)   (find and rediscover)

Why this order? The indexer uses frontmatter description preferentially for embedding. Descriptions first → better search and Discover quality.

Minimal setup: Skip step 1, just Rebuild and search.
Best quality: Generate descriptions (preview) → review → Apply → Rebuild index.

Discover Workflow

The Discover tab has two modes:

Current note — Shows notes related to whatever you're reading. Cold notes are highlighted — these are your blind spots.
Global — Shows Cold notes most related to your entire Hot pool. Great for finding forgotten gems after importing lots of files.

Click Generate MOC to export results as a linked note.

Settings

Search & Index

Setting	Default	Description
Server URL	`http://localhost:11434`	Ollama or OpenAI-compatible server
API format	Ollama	Ollama or OpenAI-compatible
API Key	—	Optional, for authenticated servers
Embedding model	`qwen3-embedding:0.6b`	Model for vector embeddings
Top results	10	Max results in search and Discover
Min score	0.5	Similarity threshold (0–1). Lower = more results
Max embed chars	2000	Content truncation. Notes with descriptions use description instead
Hot days	90	Notes created within N days are Hot
Search scope	Hot only	Hot / All / Cold
Chunking mode	Off	Off / Smart / All
Chunk size	1000	Characters per chunk
Chunk overlap	200	Overlapping characters
Exclude patterns	`_templates/` `.trash/` `3_wiki/`	Folders to skip
Synonyms	—	`keyword = syn1, syn2` per line
Auto-index	On	Re-embed on file change. Keeps Discover fresh

Description Generator

Setting	Default	Description
LLM model	`qwen3:1.7b`	Recommended: fast, good quality
Min description length	30	Shorter descriptions get rewritten. Good descriptions improve search and Discover

Commands

All commands are prefixed with Vault Search: in the Command Palette (Cmd/Ctrl+P).

Command	Description
Semantic search (modal)	Quick search with keyboard navigation
Open search panel	Sidebar with Search and Discover tabs
Find similar notes	Related notes for current file
Discover related Cold notes	Global discover — find hidden gems
Rebuild index	Full re-index
Update index	Incremental update
Generate descriptions (preview)	LLM generates descriptions → report
Apply descriptions	Write previewed descriptions to frontmatter

How It Works

┌─────────────┐     ┌──────────┐     ┌──────────────┐
│  Your Notes │────▶│  Ollama  │────▶│ Vector Index │
│  (.md)      │     │ Embed API│     │ (index.json) │
└─────────────┘     └──────────┘     └──────┬───────┘
                                            │
┌─────────────┐     ┌──────────┐            │
│  Your Query │────▶│  Ollama  │──── cosine similarity
│             │     │ Embed API│            │
└─────────────┘     └──────────┘     ┌──────▼───────┐
                                     │   Results    │
                                     │ (ranked)     │
                                     └──────┬───────┘
                                            │
                               ┌────────────▼────────────┐
                               │   Discover (no Ollama)  │
                               │   Pure vector math on   │
                               │   existing embeddings   │
                               └─────────────────────────┘

Index — Note content (or description) → embedding model → vector stored in index.json
Search — Query (+ synonym expansion) → same model → cosine similarity → ranked results
Discover — No API calls. Compares existing embeddings to surface related Cold notes
Hot/Cold — Linked/recent = Hot. Orphan = Cold. Discover highlights Cold notes in your blind spot
MOC — Export results as a Map of Content note with wikilinks and previews
Descriptions — Local LLM summarizes notes → stored in frontmatter → used for better embeddings

Recommended Models

Model	Size	Use	Notes
`qwen3-embedding:0.6b`	639MB	Embedding	Best for Chinese + English
`nomic-embed-text`	274MB	Embedding	Lighter, English-focused
`qwen3:1.7b`	1.4GB	LLM	Good quality, handles 2000+ chars
`gemma3:1b`	815MB	LLM	Lighter, but unstable > 500 chars input

For 8GB RAM machines, use qwen3-embedding:0.6b + qwen3:1.7b. Both fit comfortably.

Development

git clone https://github.com/notoriouslab/vault-search.git
cd vault-search
npm install
npm run dev    # watch mode
npm run build  # production build

License

MIT