Repository navigation
Private AI: a cloud provider you trust can be allowed for private notes. Plus effort for Claude, the date by slash
command, and fixes for tabs, tables and lists.
Upgrade notes
- From 0.2.3 or newer: replace the image and restart. No database change.
- From 0.2.2 or older: the database is updated on start (v24, see 0.2.3) – take a database backup first.
- Release rules “only for AI in your own network” keep their meaning: they are now called “Allowed – private AI
only”, and every provider in your own network is ticked as private. A cloud provider sees them only after you tick
“Allow for private notes” on it.
Added
- Private AI: each AI provider has the tick Allow for private notes (on by itself for an AI in your own network).
Notes, folders and sources allowed for “private AI only” always go to a private AI – summarise, continue and ask –,
whichever provider is the default; with several, admins choose Default private and every user can pick My AI
for private notes. A provider chosen explicitly without the tick is refused, nothing is sent. - Asking over mixed areas: the default provider answers without the private notes and says how many matching ones
it left out, with a button “Ask with …” that asks the private AI in the same chat. A chat that holds private
content then stays with private AI. - Effort for Anthropic: Claude models get Effort (model default, low, medium, high, very high) instead of
temperature; only the levels the model supports are offered. - Date and time: the slash commands
/date,/timeand “Date and time” insert the current date and time as text,
in the format of the account's language. - All tabs: with many tabs open, a button “All tabs (n)” lists every tab to switch to or close.
Changed
- Many tabs get narrower first and then scroll sideways (also with the mouse wheel) – without a scroll bar covering
the tab names.
Fixed
- Anthropic answered with an error (400, “temperature is deprecated”) on current models; Kaveyl Wiki no longer sends
a temperature there. - Tables no longer run under the table of contents: long cell text wraps.
- The folder row in the sidebar is marked when the folder's page is open in the active tab.
- Lists: an empty item in a checklist no longer turns the whole checklist into a raw block when saved, and two
lists directly after each other stay two lists (the second gets another marker). - A follow-up question in the same chat shows its answer at once (it appeared only after a while).
- “Test connection” for Anthropic lists the models even when the effort levels of the entered model cannot be read; a
level the model does not know is not kept. - Linux: a middle click on a row in the sidebar or on a tab (open in a new tab) no longer pastes the last marked
text into the open note.
Security
- Chats from before 0.2.14 in which an AI in your own network answered stay with private AI: a follow-up no longer
sends their history to a cloud provider. - An AI that may see nothing in a source no longer learns the source's or folder's name from the question's context, a
source someone cannot see is refused as the context of a question, and someone who may not read a place no longer
learns its release rule from an error message. - Swapping list markers took quadratic time for very long lists (only in the editor; the server's limits stayed).