Prism 1.4.0 rebuilt Ask so that a citation either exists or the answer does not claim it. Three point releases followed, all about the app telling you what happened.
How Ask retrieves and cites sources
The old Ask copied the first eight results off the screen, trimmed each one, pasted the block into an ordinary chat message and asked the model to name its sources. The citations that came back were prose. A model could name the wrong document or claim support that was not in the text, and nothing in the pipeline could tell the difference.
Ask now runs retrieval and cites what it retrieved. The sources under an answer come from the search itself, not the model’s account of it. You no longer need to run a search first: type a question and it goes looking.
Weak retrieval used to produce the same confident answer as good retrieval. Now an answer built on thin results is labelled as such, and when nothing matches Prism stops and says so, naming the sources it searched. You can still ask for a model-only answer. It arrives clearly marked as one.
Each connector reports its own outcome as the search runs: a result count, a quiet zero, an error or a timeout. Before, a source with no matches and a source refusing connections both showed up as nothing at all, so one dead host made the whole search look broken. Failed sources say why when you click them, and screen readers hear each source settle.
Ranking that learns, locally
Opening or copying a result is now a signal. Results you have chosen before rank higher for the same query, decaying over about a month so old habits fade. The record is a local table and nothing about it leaves the device.
The palette opens faster
Startup used to parse roughly 2 MB of JavaScript before the first keystroke, including a syntax highlighter, a Markdown renderer and two document-export libraries. That is now 487 KB, with the heavy pieces loading on the first answer, the first code block and the first export.
Fixes in 1.4.1 to 1.4.3
1.4.1 fixed a launch that showed nothing at all. Prism read every saved connector credential from the keychain before drawing anything, and macOS sometimes stops to ask for your keychain password. That question was being asked before there was a window to ask it in. Credentials now load in the background, the window comes up first, and an unanswered prompt times out with a list of the sources it could not reach.
A hotkey that fails to register now says so, in a startup notice and beside the shortcut field, and names the likely culprit. macOS gives Cmd+Option+Space to Finder search by default.
1.4.2 fixed settings that looked saved and were not. If you pointed embeddings at your own provider and pasted a key, Settings accepted the provider, the endpoint and the key and kept none of them. Search looked fine because it read the setting off the screen. Indexing kept using the local LM Studio address with no credential. All three now reach the indexer, and the key goes to your OS keychain.
1.4.3 stopped Prism blaming macOS for a service that is down. A refused connection means your machine reached something and it answered, so there was nothing to unblock. Timeouts and silent drops, which genuinely do look like the macOS network block, still name it.
Prism runs at prismapp.xyz.