Triage support tickets on sample data

Scenario

A fictional support desk receives 300 ticket threads a week across billing, delivery, and account topics. Your job: classify every thread, route each to its queue, and deliver the routed queue downstream. Sample threads included — including three deliberately ambiguous ones.

Image: sample thread inbox with topics unassigned.

Definition of done

• Every thread classified with a topic and a confidence you can read.
• Each topic routed to its queue; the routed queue delivered to the destination.
• The three ambiguous threads handled deliberately (corrected or rejected with reasons), not silently accepted.

Steps

1. Connect the sample thread source. Same catalog flow as Connect data — note the document shape of threads versus the rows you may be used to.
2. Describe the routing job. "Classify each thread as billing, delivery, or account, route by topic, and write the routed queue for the support leads." The proposal should show extraction, then classification, then routing, then a destination write.
3. Check the extraction first. Open the sample outputs: which fields came out of the threads (order numbers, dates, sentiment cues)? Bad extraction here becomes bad classification everywhere — fix field rules before touching topics.

Image: extracted fields beside their source thread, with one mis-extracted date flagged.

4. Read the topic definitions in the proposal. Billing versus account is where the ambiguity lives; confirm the boundary matches how your team actually triages before running.
5. Dry-run and hunt the three ambiguous threads. They are built to straddle topics. For each: does the confidence reflect the ambiguity? A confident wrong label is worse than an honest uncertain one.

Image: dry-run preview with confidence bars, the three ambiguous threads sitting mid-scale between topics.

6. Commit with an approval on the destination write. The routed queue leaves Semogram — gate it at the node. Review the pending approval with the per-topic counts attached, then approve once, explicitly.
7. Verify downstream and review. Open the delivered queue, confirm counts per topic, then correct or reject the ambiguous three with reasons. Your corrections teach the classifier, not just this run.

Image: routed queue downstream with per-topic counts and the three flagged ambiguous threads.

What good looks like

• Clear threads routed confidently and correctly.
• Ambiguous threads visibly handled, each with a recorded decision.
• The rerun shows the corrections took effect.

Next

• Classification depth: Forecasts.
• Approval-gated writes: Write policies.
• Same flow, your data: Run your first pipeline.