how suggestions work
miniti checks the latest finalized transcript text locally for phrases such as “could we build this?”, “look into why this fails”, or “is there a way to automate this?”. when it finds one, it can show a dismissible worth investigating prompt. this check is a small on-device phrase rule. it does not call an ai model, search the web, or use api quota. you can also use investigate at any time without waiting for a suggestion.research the web
available on macos and ios:- start a meeting and wait for some transcript text
- choose investigate
- on macos, choose research the web. on ios, the button starts web research directly
- keep talking while openai researches the recent question
- review the answer and open any cited source links inside the result sheet
investigate a codebase on macos
- open settings → ai & models → meeting investigations
- choose one codebase folder
- during a meeting, choose investigate → investigate codebase
- skips hidden, generated, dependency, build, and symlinked content
- reads no more than 1,200 supported source or documentation files, up to 300 kb each
- sends excerpts from no more than 8 ranked files, capped at 40,000 characters total
- redacts likely credential assignments and private-key blocks before sending
what is sent
an investigation can include:- the question or statement you chose
- meeting title
- current meeting notes, including locally stored calendar prep notes carried into the meeting
- up to 14 recent finalized transcript turns
- for a mac codebase investigation, the bounded local excerpts described above
gpt-5.4-mini-2026-03-17. they do not consume transcription minutes, but byok users pay openai for their own requests.
related
- google calendar - add prep notes before a meeting
- ai insights - the normal live and saved insight views
- byok - use your own openai account

