Vault Search

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

繁體中文

---

Vault Search focuses on making "search by meaning" dead simple and genuinely useful.

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. We believe your original notes have intrinsic value. The best search system doesn't replace your writing — it helps you rediscover it. RAG and semantic search shine precisely because they work with your existing content, not over it.

What sets Vault Search apart

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

Simple and fast — Sidebar panel persists results while you browse. Cmd/Ctrl+P for instant modal search. One-click "Find Similar" with zero API calls. Clean UI, intuitive workflow.

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.

Smart prioritization — Hot/Cold tiers automatically surface linked and recent notes, keeping search results relevant to your actual workflow. Cold (orphan) notes stay searchable but don't dilute results.

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 relevance for long notes — a feature rarely seen in lightweight plugins.

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.

Optional chunking — For users with long documents, enable chunking to search within specific sections. Three modes:

• Off (default)
• Smart (skip notes with descriptions)
• All

Most users don't need chunking, but power users can adjust as needed.

"Strikes an ideal balance between privacy, simplicity, Chinese language support, and search quality."

Features

• Semantic Search — Find notes by meaning, not just keywords. The more you describe, the more precise the results.
• Sidebar Panel — Persistent results in the right sidebar
• 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 — Prioritize linked and recent notes
• Chunking — Optional overlapping chunks for long documents (off by default)
• Description Generator — Local LLM generates frontmatter descriptions for better search quality
• Synonym Expansion — Define synonyms to improve recall
• Multi-format API — Ollama + OpenAI-compatible (LM Studio, llama.cpp, vLLM)
• Bilingual UI — English & Traditional Chinese

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 (no mobile support planned yet)

Installation

BRAT (recommended)

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

Manual

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

Quick Start

1. Settings → Vault Search → Select your embedding model
2. Click Rebuild to index your vault
3. Cmd/Ctrl+P → "Semantic search" or click the ribbon icon

Recommended Workflow

For the best search quality, follow this order:

1. Generate descriptions  →  2. Rebuild index  →  3. Search
   (LLM summarizes notes)    (embed with descriptions)   (enjoy better results)

Why this order? The indexer uses frontmatter description preferentially for embedding. If you generate descriptions first, the index captures high-quality summaries instead of raw content — resulting in noticeably better search relevance, especially for long notes.

• Minimal setup: Skip step 1, just Rebuild and search. Works fine for short notes.
• Best quality: Run Generate descriptions (preview) → review the report → Apply descriptions → then Rebuild index.
• With chunking: Use "Smart" mode — notes with descriptions use the description for embedding; notes without get chunked automatically.

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 shown
Min score	0.5	Cosine similarity threshold (0–1)
Max embed chars	2000	Content truncation for embedding
Hot days	90	Recent notes threshold
Search scope	Hot only	Hot only or all notes
Chunking mode	Off	Off / Smart / All
Chunk size	1000	Characters per chunk
Chunk overlap	200	Overlapping characters
Exclude patterns	_templates/ .trash/ .obsidian/	Folders to skip
Synonyms	—	keyword = syn1, syn2 per line
Auto-index	On	Re-embed on file change

Description Generator

Setting	Default	Description
LLM model	qwen3:1.7b	Model for generating descriptions
Min description length	30	Shorter descriptions get rewritten

Commands

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

Command	Description
Vault Search: Semantic search (modal)	Quick search with keyboard navigation
Vault Search: Open search panel	Sidebar with persistent results
Vault Search: Find similar notes	Related notes for current file
Vault Search: Rebuild index	Full re-index
Vault Search: Update index	Incremental update
Vault Search: Generate descriptions (preview)	LLM generates descriptions → report
Vault Search: Apply descriptions	Write previewed descriptions to frontmatter

How It Works

┌─────────────┐     ┌──────────┐     ┌──────────────┐
│  Your Notes │────▶│  Ollama  │────▶│ Vector Index │
│  (.md)      │     │ Embed API│     │ (plugin data)│
└─────────────┘     └──────────┘     └──────┬───────┘
                                            │
┌─────────────┐     ┌──────────┐            │
│  Your Query │────▶│  Ollama  │──── cosine similarity
│             │     │ Embed API│            │
└─────────────┘     └──────────┘     ┌──────▼───────┐
                                     │   Results    │
                                     │ (ranked)     │
                                     └──────────────┘

1. Index — Note content (or description if available) → embedding model → vector stored locally
2. Search — Query (+ synonym expansion) → same model → cosine similarity → ranked results
3. Hot/Cold — Linked/recent = hot (default). Orphan = cold (opt-in)
4. Chunking — Long notes split into overlapping chunks, each embedded separately. Search returns the best-matching chunk's score.
5. Descriptions — Local LLM summarizes notes → stored in frontmatter → used preferentially for embedding

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