Palette and search

The Palette is onQ's fastest path from intent to clipboard. It searches the open vault and also exposes creation, lock, similarity, and plugin commands.

Open the Palette

The default global shortcut is:

Platform	Default
Windows	Win+Q
Linux	Meta+Q
macOS	⌘+Q

The shortcut restores the window from the tray and toggles the Palette. Inside the app, Ctrl+Q or Cmd+Q also toggles it. Record a different chord from Settings > General > Global shortcut.

Empty-query behavior

With no search text, the Palette shows:

1. New prompt
2. Lock vault
3. enabled plugin commands, when any exist
4. up to five recently copied or edited prompts
5. current similarity results, when a More like this lookup has run

Recent prompt IDs live in the local WebView profile. If no recent history exists, onQ falls back to the five most recently updated vault prompts.

Free-text search

Typing starts a debounced search after 150 ms. The searchable document for each unlocked prompt combines:

• title;
• tags;
• project path;
• body.

The backend returns at most 50 results for the Palette. Results include title, metadata, and a short body snippet. A locked result never includes a plaintext body snippet.

The Palette input is free text. It does not parse tag:, folder:, recent:, or lock: operators. Structured operators are only accepted by the smart-folder DSL.

Keyword retrieval

Keyword retrieval is always available. onQ:

1. splits text on non-alphanumeric characters;
2. lowercases each token;
3. hashes tokens into a deterministic sparse vector;
4. normalizes term frequencies so long prompts do not win only by length;
5. queries MongrelDB's sparse index.

This is token matching, not substring matching. Search for meaningful words from a title, tag, project, or body.

Optional semantic retrieval

Semantic search uses sentence-transformers/all-MiniLM-L6-v2 through ONNX Runtime:

• output dimension: 384;
• pooling: attention-mask-aware mean pooling;
• output normalization: L2;
• inference: local;
• download source: Hugging Face Hub;
• approximate model size: 90 MB.

The model is not downloaded on first launch. Select Settings > Search > Load MiniLM model. onQ downloads onnx/model.onnx and tokenizer.json into the platform app cache, loads the model, and re-embeds every prompt.

If the model is absent or fails to load, search continues in keyword-only mode. Existing prompt creation and editing do not need network access.

Ranking

When semantic search is available, onQ queries both:

• sparse keyword relevance;
• approximate nearest-neighbor vector similarity.

It combines emitted result ranks with Reciprocal Rank Fusion. Two small tie-breakers then favor:

• recently updated prompts;
• favorites.

The Search recency half-life setting controls the exponential recency decay period. The default is 30 days. Larger values preserve the recency boost for longer; smaller values make it disappear faster.

Favorites receive only a small bonus. A strongly relevant non-favorite should still outrank a weak favorite.

Embedding index modes

After the model is loaded, Settings > Search offers:

Mode	Behavior	Use when
Binary, default	Binary HNSW candidates followed by exact cosine reranking	Lower memory and fast retrieval
Dense	Full-precision dense ANN after index publication	Recall matters more than memory

Changing mode persists the preference and rebuilds the derived embedding index. Dense mode can use exact cosine temporarily while the replacement index is pending.

Neither mode changes prompt files. The embedding index can be rebuilt.

More like this

Open an existing prompt in the editor, then open the Palette again. The More like this: title command becomes available.

This path uses MinHash over character trigrams, not MiniLM. It works without the optional model and returns up to ten prompts with similar body shingles. The source prompt is removed from its own results.

Smart-folder DSL

Smart folders accept structured filters:

Token	Meaning	Example
text:<value>	Add free text to ranked retrieval	text:"release notes"
folder:<path>	Exact project-path match	folder:"Work/Releases"
tag:<tag>	Add a tag to the tag-any filter	tag:writing
-tag:<tag>	Remove that tag from tags already named in this DSL	-tag:draft
favorite:true or false	Filter favorite state	favorite:true
locked:true or false	Filter lock state	locked:false
char:>N	Minimum character count	char:>500
char:<N	Maximum character count	char:<5000

Values with spaces may use single or double quotes. Multiple text: values join with spaces. Multiple positive tag: values use any-tag matching, not all-tag matching.

-tag: is not a general exclusion operator. It only removes a matching positive tag token from the query being assembled. For example, tag:draft -tag:draft leaves no tag filter.

Unknown operators are rejected. The DSL cannot run commands or scripts.

Examples:

favorite:true tag:writing

text:"incident response" folder:Operations locked:false

char:>200 char:<4000

The visual builder can create tag, folder, favorite, locked, and text predicates. Character bounds require direct DSL editing.

Search and locked prompts

Locking a prompt removes its body from the Markdown file and sparse index. Title, project, tags, favorite state, timestamps, and the lock marker remain visible. A previously generated semantic embedding is preserved in the protected search index, so the prompt may still rank semantically while locked.

Copy and edit actions refuse to expose the locked body. Unlock the prompt from its editor first.

Relevance tips

• Put the words you normally recall in the title.
• Use a few stable tags rather than many near-duplicates.
• Use project paths for ownership or workflow, not every topic.
• Favorite high-value prompts.
• Load MiniLM only when semantic recall is useful enough to justify model storage and re-embedding time.
• Use smart folders for exact metadata filters. Keep Palette input as ordinary search text.

Diagnose search

Open Settings > Search and inspect:

• keyword index status;
• model loaded, cached, or not installed;
• sparse-only or semantic path;
• binary/dense mode and readiness.

Use Refresh status after changing model files or waiting for a rebuild. See Troubleshooting when status does not change.