Purpose
Build the candidate person set used by speaker verification, follow-up drafting, and relationship context.
When to use it
- The transcript mentions people not present on the invite.
- The calendar roster is incomplete or unavailable.
- You need to distinguish attendees, mentioned people, and owners.
Inputs
- Transcript segments
- Calendar roster
- Known people list
Outputs
- Candidate attendees
- Mentioned non-attendees
- Unresolved person references
Steps
- Collect names and email handles from calendar, transcript, and source metadata.
- Classify each person as expected, spoken, mentioned, or unresolved.
- Merge duplicates only when evidence is strong.
- Return unresolved nicknames and pronouns for review.
Quality checks
- Attendee status is not treated as speaking proof.
- External people and internal people are not merged by first name alone.
Failure modes
- A common first name maps to the wrong person.
- Transcript labels use roles instead of names.
Privacy notes
- Person rosters can reveal sensitive customer or employee context.
- Keep unresolved names out of public artifacts.
Example prompt
Extract candidate attendees and mentioned non-attendees. Separate expected attendance from evidence that someone spoke.
Example structured output
{"expected":["Dana","Lee"],"spokenCandidates":["Dana","Morgan"],"mentionedOnly":["Erich"],"unresolved":["Ray?"]}