Apple Notes — Semantic Search & RAG

Search 7,500+ Apple Notes by meaning and ask questions answered from note content. Entirely local — nothing leaves the machine.

notes is on PATH (wrapper at /usr/local/bin/notes). No venv activation needed.

Ollama must be running (nomic-embed-text for embeddings, qwen3.5 for RAG).

Prefer `search` over `ask`

Start with search. It's fast (0.1s), gives you titles + scores + snippets, and is enough for most questions. Only escalate to ask when the user needs a synthesized answer across multiple notes.

Many notes are short bookmarks (a title + a link). search shows you exactly what's there. ask adds 10–30 seconds of LLM time and may just report "I only see titles" for thin notes.

Commands

Search notes by meaning

notes search "query" [--limit N] [--folder FOLDER] [--json]

Returns the top N notes ranked by similarity (0–1, higher = better). Default: 10.

--folder "Embodied Labs" — filter to a specific folder
--json returns: [{id, title, folder, body, score, modified}, ...]

Examples:

notes search "restaurant recommendations"
notes search "visionOS development" --limit 20
notes search "budget expenses" --json | jq '.[].title'
notes search "cast members" --folder "Embodied Labs"

Ask a question answered from notes

notes ask "question" [--context N] [--folder FOLDER] [--model MODEL] [--json]

Retrieves relevant notes, feeds them to qwen3.5 locally, streams an answer with citations. Default: 10 context notes.

--json returns: {answer, sources: [{id, title, score}]}
--context 20 gives the LLM more notes to work with

Examples:

notes ask "What machine learning resources have I saved?" --context 20
notes ask "Summarize my notes about home automation"

Open a note in Apple Notes

notes open <id>

Takes a note ID (from search results) and opens it directly in Apple Notes.app. Useful for viewing attachments, images, and link previews that aren't in the text.

Example workflow:

notes search "framework for creating skills" --json | jq '.[0].id'
# → 14145
notes open 14145

Sync the index

notes sync [--force]

Incremental — only re-embeds changed notes. ~0.3s if nothing changed, ~40s for full rebuild.

Run sync before searching if the user says they recently added or edited notes.

Check index status

notes status

Interpreting results

Scores 0.85+: Strong semantic match.
Scores 0.70–0.85: Relevant, worth showing.
Below 0.70: Weak — probably not what the user wants.
55% of notes contain URLs (mostly X/Twitter posts, GitHub repos, YouTube). The URL is in the body text and searchable.
Many notes are title + link bookmarks: The user saves links from X, GitHub, etc. with short descriptive titles. Search works on both the title and URL text.
If a note body is very short (just a title), the note likely has an image/screenshot/link-preview attachment. Suggest notes open <id> to view it in Apple Notes.
Folders: Notes (7,574), Embodied Labs (6), IRS (1), GPJ (1). Almost everything is in "Notes".

When to use which

User intent	Command	Why
"Find notes about X"	`search`	Fast, shows what exists
"Do I have anything about X?"	`search --limit 5`	Quick existence check
"What did I write about X?"	`search` first, `ask` if notes have real content	Avoid slow RAG for title-only notes
"Summarize my notes on X"	`ask --context 20`	Needs synthesis across notes
"Open that note"	`open <id>`	View in Apple Notes with attachments
Pipe to another tool	`search --json` or `ask --json`	Structured output

Important

All local: Ollama on localhost. No notes leave the machine.
Search is meaning-based, not keyword. "dining spots" finds restaurant notes.
Short notes (< 5 words) may cluster together in search — the embedding model needs context to differentiate very short texts.
First ask after idle may take 30–60s — Ollama cold-loads qwen3.5 into GPU memory.
Source code: /Users/dalecarman/Groove Jones Dropbox/Dale Carman/Projects/dev/apple-notes-export/

apple-notes