ticket-category-analysis

Export tickets with Ticket ID, Month, and Description columns, then run prepare_batches.py to deduplicate descriptions into cases. Classify each batch into the taxonomy paths, and apply_classification.py validates every path and scores accuracy against an answer key if one exists.

Run analyze.py on the categorized workbook to produce a self-contained interactive HTML dashboard, then build_deck.py to generate a nine-slide PowerPoint. Both draw numbers from the same metrics module so the deck cannot disagree with the dashboard.

Yes, any helpdesk export works as long as it contains Ticket ID, Month, and Description columns. If headers differ, pass a column map like --column-map 'number=Ticket ID' and optional columns such as Priority or SLA Met simply unlock extra report sections.

The pipeline reports that accuracy was not measured rather than pretending it passed. Hand-label 200 to 300 tickets into an Answer Key sheet with true L1, L2, and L3 values, then rerun to get a real accuracy number per category and confidence band.

Assignment Group usually maps one-to-one onto the category because the desk already routed the ticket, so feeding it to the classifier means reading the answer off the routing. The resulting accuracy score measures nothing about text-based categorization.

A subcategory must clear three explicit checks: at least 12 tickets in the latest month, resolution-time spread of 3 hours or less, and no more than 25% P1 or P2 tickets. Thresholds are overridable via flags, and the report shows actual values beside each check.