Nightly duplicate-ticket clustering with a human review queue

Runs a batch clustering pass over open tickets and surfaces likely duplicate groups for a human to confirm, instead of merging automatically. Lower-risk than real-time auto-merge, it catches duplicates that share a topic but not exact wording.

Use it when your team wants the speed of clustering but not blind auto-merges — for example, regulated support desks or teams burned by false-positive merges. The nightly cadence keeps the morning queue deduplicated.

1. 1A scheduled trigger runs overnight.
2. 2Zendesk returns all tickets still open from the last 24 hours.
3. 3OpenAI embeds each ticket's text.
4. 4A logic step runs agglomerative clustering on the embeddings, grouping tickets above a similarity cutoff into candidate clusters of two or more.
5. 5Clusters are written to Postgres with a pending status so approvals survive restarts.
6. 6Each cluster is posted to a Slack review channel with ticket summaries and Approve / Reject buttons; an approval triggers the merge back in Zendesk.