What you'll have at the end

A notebook that groups your tickets into topic clusters and names each one, ready to skim in an afternoon.

Before you start

You need

A batch of raw support ticket text already pulled from your helpdesk, plus a working call to an embedding model that turns one piece of text into one vector.

Not covered

Reclassifying new tickets automatically as they arrive; this recipe sorts one batch already sitting in front of you rather than running as a live pipeline.

Leans on

Embeddings Cheat SheetIts pooling and normalization tables cover how a batch call turns into one vector per ticket, and whether that vector needs a length fix before you measure distance.
Unsupervised Learning Cheat SheetIts clustering algorithm and validation metric tables carry the two choices this recipe leans on: which grouping method to run, and how to check the groups it finds are real.

Where to go next

Verify your first embedding works

go there first if you don't yet have a working embedding call to turn ticket text into vectors

Choose cosine similarity or dot product for your vectors

go there for the fuller reasoning behind matching your distance metric to your vectors; this recipe only applies that choice

Untangle five different phrasings of the same question in your dataset

go there instead if you're deduplicating near-identical repeats of one question rather than sorting a whole backlog into named themes

Support tickets sorted into topics without hand labeling every one

Before you start

Where to go next

Support tickets sorted into topics without hand labeling every one

Before you start

Where to go next