Lina

AI-powered note assistant and hybrid search engine for Obsidian. Features local indexing, semantic search, contextual slash commands, and privacy-first AI analysis.

User Manual | Commands Guide | Changelog

Overview & Key Features

Lina helps you find, connect, and enrich Markdown notes in Obsidian without taking control away from you.

• 🔍 Hybrid Search: Combines fast local text indexing with semantic vector embeddings into a single ranked list.
• ⚡ Contextual Slash Commands: Execute /ask (AI note query), /tags (smart tag suggestions), and /yaml (frontmatter generation) directly from the sidebar input.
• 🔒 Privacy First: All indexing data is stored locally in .lina/index/. Zero network requests by default. Remote AI providers are contacted only when explicitly configured and triggered.
• 🛡️ Explicit Confirmation: AI responses, tag additions, and YAML fields are applied to active notes only after your explicit confirmation.
• 📱 Mobile & Sync Friendly: Memory-aware safeguards for mobile devices (validated on Android) with opt-in binary shadow sets and Syncthing support.

Quickstart & Installation

Option 1: Manual Installation

1. Download manifest.json, main.js, and styles.css from the latest release.
2. Create the directory <Vault>/<configDir>/plugins/lina/ (where <configDir> is your vault's Obsidian config directory, default .obsidian).
3. Copy the downloaded files into that folder.
4. Enable Lina under Obsidian Settings > Community Plugins.

Option 2: Community Plugins

Search for Lina in Obsidian Community Plugins once officially listed.

First-Time Indexing Note:
When installed for the first time, Lina does not auto-build the full index. Open the Lina side panel and click Rebuild Index (or run the rebuild index command). Once a valid index exists, Lina keeps it updated automatically in the background.

Features & Side Panel

The Lina panel lives in Obsidian's right sidebar. It serves as your search interface and AI assistant hub.

Search Modes

• Hybrid (Recommended): Combines text (default weight 0.7) and semantic similarity (default weight 0.3).
• Text: Fast local search by note title, path, or content. Supports exact, prefix, and substring matching.
• Semantic: Meaning-based search powered by vector embeddings. Requires generated embeddings.

Contextual Slash Commands (/ask, /tags, /yaml)

Type a slash command in the sidebar input to interact with your notes using AI:

• /ask <prompt>: Queries the configured AI provider about the active note context. Insert responses below selection, replace selection, or append to note after confirmation.
• /tags: Asks AI to suggest tags for the context. Displays checkboxes to apply non-duplicate tags with confirmation.
• /yaml: Asks AI to suggest frontmatter fields. Displays checkboxes to safely apply new fields without overwriting existing data.

Context Selection Order:

1. Selected text in active Markdown editor.
2. Preserved selection captured from the active note (if focus shifted to panel).
3. Active note content.

[!NOTE] All slash commands recheck context against configured path and term exclusions before contacting an AI provider.

AI Providers & Model Configuration

Lina allows separate provider and model configurations for Analysis AI (Chat/LLM) and Embeddings.

Configuration Details

• Base URLs: Automatically populated for Ollama (http://localhost:11434), Mistral (https://api.mistral.ai/v1), and OpenRouter (https://openrouter.ai/api/v1), customizable.
• Model Selection: Ollama and Mistral select models via dropdown catalog; OpenRouter uses a free-text input field for model identifiers.
• API Keys: Per-device structure. Keys start empty and require explicit save or clear actions.
• Batch Size: Configurable (1–50) for native batching with Mistral and modern Ollama (/api/embed). Legacy Ollama endpoint fallback processes 1 item per request.

Mobile & Multi-Device (Syncthing)

Desktop & Mobile Support

• isDesktopOnly: false. Manually validated on Desktop (Windows) and Android (Samsung Galaxy S23 Ultra, One UI 8.5, 8 GB RAM).
• On mobile devices, local Ollama is usually unavailable. Remote providers (or text-only search) are recommended.
• Memory safeguards prevent dangerous allocations. If embeddings exceed safe memory budgets, Lina reports no-safe-source and falls back to text search.

Recommended Syncthing Workflow ("PC Producer / Mobile Consumer")

1. PC Producer: Generate the text index and canonical embeddings.jsonl on desktop. Optionally maintain the experimental binary shadow copy (embeddings.vectors.f32).
2. Sync Vault: Sync the .lina/index/ directory to your mobile device via Syncthing.
3. Mobile Consumer: Mobile syncs the pre-built index and validates the publication id, using the safe index for search without high-memory rebuilding overhead.

For full .stignore rules and step-by-step instructions, see the Syncthing Guide in the User Manual.

Experimental Binary Embedding Storage

An opt-in derived shadow copy (embeddings.binary.manifest.json, embeddings.meta.jsonl, embeddings.vectors.f32) is available for advanced users:

• Canonical Source: .lina/index/embeddings.jsonl remains the source of truth.
• Preference: maintainBinaryEmbeddingCopy (default off) and embeddingStorageReadPreference (prefer-binary).
• Safety: Binary is accepted only when all three files are valid and match the canonical publicationId. Falls back safely to JSONL or text search.

Privacy & Local Data Storage

• Data Path: Local index operational data is stored strictly in .lina/index/.
• No Web Storage: Lina does not use localStorage or sessionStorage. Settings use Obsidian's loadData/saveData APIs.
• Network Boundaries: Zero network traffic by default. External APIs are called only when you explicitly configure a remote provider and run an action.

Development

# Install dependencies
npm ci

# Build plugin bundle
npm run build

# Run official Obsidian linters
npm run lint:obsidian
npm run lint:obsidian:strict

Key build artifacts: manifest.json, main.js, styles.css.

License

MIT License