“What did the customer say about pricing in the last three calls?” is exactly the kind of question a language model is good at answering. It is also exactly the kind of answer you cannot trust without checking, because a fluent paragraph with a made-up timestamp looks identical to a correct one.
unottr’s Ask and overview features were designed around that problem. The rule is short: the model is not allowed to invent a citation, and a citation that does not resolve is dropped rather than shown.
How a citation works
Before the model sees anything, unottr retrieves the transcript segments, fresh overviews and current tasks that match your question. Each segment is handed over as [id] Name: text, and the model is told to cite ids.
Ids resolve to milliseconds in one place. When an answer comes back, each citation is checked against the set that was actually retrieved. A valid one shows the recording, speaker, date, timestamp and the current excerpt, and opens the video at that second. An unknown one is removed. The model cannot produce a confident link into the wrong minute, because the only links it can produce are to text it was shown.
The same mechanism drives AI overviews: a title, a TL;DR, topic sections, decisions and a task list, with every line clickable back to the moment it came from.
Scope you can see, and that does not drift
An Ask thread starts with a scope: particular recordings, people or a range of meeting dates. You see it before you send. Once the thread starts, the scope is pinned, so a follow-up question cannot quietly search a different set of meetings than the one you were looking at.
Threads stay in the local database. They can be searched, renamed and deleted like anything else in the library.
Regenerate without losing your edits
Overviews are regenerable, and regeneration is where most summary tools throw away your work. In unottr, prose is overwritten but tasks you have edited, checked off or dismissed survive. If you corrected a transcript after an overview was written, the overview says its sources changed instead of pretending to be current.
Bring your own model
None of this requires a particular AI provider. Ask works with Ollama or LM Studio on your own computer, with OpenAI, Anthropic, Mistral, OpenRouter or any OpenAI-compatible endpoint, or with an installed Claude Code, Codex or Antigravity command-line agent. The app includes the connections, not the model usage, and you pay that provider directly.
An answer is still only a starting point. The difference is that in unottr every sentence of it comes with a way to listen to the original.
