Settings¶
Open Settings from the toolbar on iPhone (the sliders button), from the bottom of the sidebar on iPad, or with Cmd+, on a Mac. Changes apply to your next question. The app remembers them.
Beside the Search, Answers and Synthesis headings is a question-mark button that opens the matching page of the in-app Help.
Corpus¶
Scope limits what a question searches. See Choose where to search. On Mac and iPad it is in the sidebar as well.
Search¶
| Setting | What it does |
|---|---|
| Results | How many passages the search returns, 1 to 50. The model reads as many as fit, best first. More results give you more to read under the answer, not a longer answer. |
| Min score | Discards matches weaker than this, 0 to 0.9. Raise it when answers draw on passages that only share a word with your question. Lower it when a search comes back empty. |
| Semantic floor | The least meaning-based similarity that a match found by exact wording must also have, 0 to 0.9. It stops a shared rare word from being enough on its own. |
Reset search to defaults puts all three back.
Answers¶
Engine picks who writes the answer; see Answer engines.
Synthesis¶
These shape how the model writes. They apply from the next answer.
| Setting | What it does |
|---|---|
| Temperature | How much the model varies its wording, 0 to 1. The default is 0.2: low, but not zero, because at exactly zero these models can loop and repeat a phrase. Above about 0.5, answers drift from the passages. |
| Greedy decoding | Always takes the most likely next word, the same as temperature 0. It greys out the temperature slider. Use it to compare two other settings fairly. |
| Instructions | Guide, the default, is written for the small on-device model and asks for two or three paragraphs. Worker parity is the wording written for the large model on a server, and gives very short answers on the on-device model. |
| Passages per work | How many passages from one book may appear in one answer. Engine default is 2 for on-device and 4 for Private Cloud. Raise it for depth in one book, lower it for variety. |
| Permissive guardrails | On-device only. Relaxes Apple's content check on your question and the answer, which classic literature trips more than you would expect. It does not turn safety off. |
Reset synthesis to defaults puts all of them back.
Illustrations¶
Resolution and Backend set the size and source of an illustration that Render draws. Illustrations always come from a worker you run; the app cannot draw them on its own. The three resolutions are Preview (768 × 512), Standard (1152 × 768) and Full (1536 × 1024). A smaller one renders faster. Backend lists only the image sources the worker reports as ready now.
Delete all conversations¶
Removes every saved chat, including their answers and illustrations. It asks first and cannot be undone.
Corpus information¶
Shows where the books are kept. With the corpus installed, it reads "on this device" and lists the vector format, the embedding model, and the size. Search and answers both run here. Airplane mode changes nothing. appears when you are on the on-device engine.
Worker¶
The address of a server you run. Type it and tap Test. The app reports the
number of books and genres it found, or why it could not reach it. On an
iPhone or iPad, localhost means the device itself, so use your Mac's name or
address instead. A GutenbergKG worker answers only its own Mac unless it is
started with WORKER_BIND=0.0.0.0 in its docker/.env; anyone on your
network can use it then. See Answer engines.
Help and About¶
Help opens the in-app Help. About (on iPhone and iPad; on a Mac it is in the app menu) shows the version and the corpus counts.