Semantic Retrieval

Semantic retrieval adds vector-based note search to the agent. It is opt-in and scoped so you control exactly what gets indexed and when.

How it works

1. Index — The plugin sends note text to an embedding provider and stores the resulting vectors in a local index file.
2. Query — When the agent needs relevant context, the index returns the most semantically similar notes.
3. Scope — Indexing is scoped to a vault, folder, tag, or project workspace. You choose the scope each time you index.

Setup

Enable and configure embeddings in Settings > Resources > Semantic retrieval:

• Provider — OpenRouter, Ollama, or OpenAI-compatible.
• Model — The embedding model id (e.g., openai/text-embedding-3-small for OpenRouter, nomic-embed-text for Ollama).
• Dimensions — Expected vector size (default 1536).
• Language coverage — Multilingual, monolingual, or unknown. Used for diagnostics.
• Batch size — Notes per embedding request (default 32).
• Max indexed characters per note — Upper bound on text sent to the provider per note (default 12,000).

Provider API keys reuse the same secrets configured in Settings > Models.

Commands

Command	What it does
/semantic-index status	Show current index state: scope, model, note count, last update.
/semantic-index estimate	Estimate how many notes and tokens an index would cover for the current scope.
/semantic-index start	Build or rebuild the index for the current scope.
/semantic-index cancel	Cancel an in-progress index build.

When a project is active, /semantic-index scopes to the project folders automatically.

Privacy

• Embeddings are sent to the provider you configure. If you use a remote provider, note content leaves your device.
• The local index file lives inside the plugin directory (semantic-index.json).
• Use Ollama for fully local embeddings.