Cluster Intercom Conversations into Recurring-Issue Themes Nightly

Turns a flood of individual Intercom conversations into a short, ranked list of recurring issue themes so your support and product teams see the real patterns instead of one-off tickets.

Run it when conversation volume is too high to read manually and you want a nightly digest of what customers actually struggled with — grouped by theme, with counts, not 400 separate threads.

1. 1A nightly schedule fires after business hours.
2. 2The flow fetches Intercom conversations closed in the last 24 hours, including transcript text and tags.
3. 3OpenAI generates an embedding for each conversation's core complaint.
4. 4A clustering step groups embeddings by cosine similarity and labels each cluster with a plain-language theme and severity.
5. 5Clusters below a minimum-size threshold are dropped as noise.
6. 6The remaining themes, with example conversation IDs and volume counts, are upserted into a Postgres `issue_themes` table keyed by theme fingerprint so recurring themes accumulate over time.