Purpose
Reduce long transcripts into inspectable clusters without losing source evidence.
When to use it
- A meeting covers multiple projects or customer issues.
- You need project detection before memory updates.
- The summary should be organized by topic rather than chronology.
Inputs
- Transcript segments
- Agenda items
- Known project names
Outputs
- Topic clusters
- Source segment references
- Project candidates
Steps
- Identify repeated nouns, project names, customers, and decision verbs.
- Group adjacent and non-adjacent segments by topic.
- Attach source segment references to each cluster.
- Label clusters with concise, reviewable names.
Quality checks
- Clusters include evidence references.
- Project candidates are marked as candidates unless matched to a source of truth.
Failure modes
- A single word means different things in different departments.
- The model over-compresses unrelated side discussions.
Privacy notes
- Topic names can expose confidential projects.
- Keep raw evidence links private when publishing sanitized summaries.
Example prompt
Cluster this transcript by durable topics. Include segment IDs and mark project candidates separately from confirmed projects.
Example structured output
{"topics":[{"name":"Order guide routing","segments":["s12","s13"],"projectCandidate":"Consolidated order guide"}]}