Themes

A dataset of real conversations holds the answer to questions no rubric asks: what are people actually using this for, where do they get stuck, and what does it keep getting wrong? Reading a thousand of them by hand isn't a plan. Themes reads them for you and groups what it finds.

It sits at the bottom of any dataset page. With no analysis yet, it offers Analyze themes.

Choosing what to look for

A theme analysis looks along dimensions — four to start with, all switched on:

• Use cases — what someone came to do, read from the whole conversation.
• User frustrations — the moments people got stuck, counted as they recur.
• Model mistakes — where the answer was wrong, counted as they recur.
• Overall sentiment — how the conversation felt, taken as a whole.

Add custom dimension takes your own. Give it a name, say whether it's about the whole conversation or recurring events, and write what to look for — being specific about what counts as an occurrence is what makes the results usable.

Two switches sit under the list:

Reading the results

Each dimension gets a grid of theme tiles: the theme's name, how many conversations it covers, and a sparkline of when those conversations happened. A tile is marked rising, steady or cooling when there's enough history to tell. Sort the grid by coverage, trend, or how recent a theme is.

Click a tile and it opens underneath:

• Also appears with — the themes that keep showing up in the same conversations. Clicking one jumps to it, which is how you find a pair that's really one problem.
• A timeline you can click. Pick a bar and the examples below narrow to that window; Range takes exact dates instead.
• The quotes themselves — the actual lines that put each conversation in this theme, each one a click away from the conversation it came from.

If a theme's name isn't the words your team uses, hover it and rename it. The name is yours; only the grouping was ours.

Keeping it current

A live analysis picks up new conversations by itself, and the page updates while you watch it. When an analysis isn't live and the dataset has grown, a Process new button appears with the count waiting, and catches it up without starting over.

Only the most recent analysis is shown. Starting a new one replaces what's on screen.

Taking it elsewhere

Export CSV downloads the themes and their counts. Export to Napkin sheet puts the same table in a Napkin sheet and opens it, which is the quicker route when the next thing you'll do is chart it or show it to someone.