AI search

Ordinary Spotlight matches characters. AI search matches meaning, so “that article about rendering on the server” can find a bookmark whose title contains none of those words.

What it needs

Three things, all of them:

1. A Pro licence.
2. Your own OpenAI or Gemini key, set up under the gear → AI Assistant. See Bringing your own key.
3. An index. Same panel: AI Search Index → Build index. It shows progress as n / total indexed.

The index is what makes search-by-meaning possible — it turns each item into a vector once, so a query can be compared against all of them. Gemini’s embeddings are free; OpenAI’s are paid, and the panel says roughly what that costs.

Change provider later and the button becomes Rebuild for openai (or gemini); the two providers’ vectors are not interchangeable.

Using it

Open Spotlight and look at the bottom right for an AI button. It only appears once a key is configured — no key, no button.

1. Click AI to arm it. Without a licence the button shows a padlock and opens the upgrade dialog instead.
2. Type a question in your own words.
3. Press Enter. Ordinary search runs as you type; AI search runs only when you ask for it.

The footer hint changes from select to search while Enter would run the query, and back once results are in — so a second Enter opens the highlighted result instead of searching again.

Reranking

The AI panel has Use LLM to rerank AI search results, on by default.

• On: the closest 30 items are handed to the model, which reorders them and writes a short reason next to each. Adds a second or three and one model call per search.
• Off: you get the raw index ranking, instantly, with no explanations.

Build the index first

With no index built, AI search does not refuse — it falls back to whatever 30 items come first and ranks those. The results look plausible and are close to meaningless. If AI search seems to ignore half your board, check the index.