v0.4.0
Live voice starts searching your Libraries while you are still speaking, so a grounded reply comes back sooner.
Released 21 July 2026. This release closes the pause between the end of your sentence and the start of a spoken reply, by searching your Libraries while you are still talking. It also adds an OpenAI-compatible API and an artefact panel; see the changelog entry for both.
Live voice searches your Libraries while you speak
In a spoken conversation, the search of your Libraries used to happen entirely after you stopped talking, sitting squarely in the gap before the reply. Live voice now searches on the part of your sentence that has already settled, so by the time you stop, the passages are usually already in hand. When your turn ends, the query that was searched for is compared with what you actually said: if you simply finished the sentence or re-worded the same question, the result is used and the reply skips its own search; if you changed your mind mid-sentence, everything speculative is discarded and the turn searches exactly as it always did. Access control is inherited rather than reimplemented: a speculative search resolves the same Library scope, with the same per-document restrictions, as the turn itself, and is audited like any other retrieval.
Upgrade
The feature is on by default and needs a streaming speech-to-text engine; with batch transcription there are no partial transcripts and it never fires. Administrators can tune how eagerly it speculates and how closely a search must match what was said before its result is used.