Encyclopedia Evalica / Observability / Topics

Topics

/'tah.pihks/An automated clustering feature that summarizes traces into simple statements, then groups them into categories like recurring issues, common tasks, or sentiment shifts. Topics help teams move from single-session debugging to "horizontal" pattern-finding. (noun)

Why it matters

Reading individual traces tells you what went wrong in a single conversation, but it cannot tell you what is going wrong across your product. Topics automate the pattern-finding step by summarizing traces into plain-language statements and clustering them into categories. This horizontal view surfaces trends that no amount of individual log reading would reveal, like a new class of user requests your system handles poorly, or a recurring failure mode that affects a small percentage of every session. Manual log review does not scale because the volume of production traces quickly outpaces what any team can read. Automated clustering lets you monitor your system at the category level, track how topic distributions shift over time, and prioritize the issues that affect the most interactions. When you spot a problematic topic, you can drill into the underlying traces, create a targeted dataset from them, and build scorers to catch that pattern going forward.

Topics surfaced a new cluster of 'invoice confusion' requests we hadn't anticipated.

Customer example

Replit uses Topics in Braintrust to surface systemic issues that wouldn't show up in individual bug reports, clustering traces into patterns the team can act on. Read more

Related Observability terms

From the docs

Get started with Evals

Braintrust is the AI observability and eval platform for production AI. By connecting evals and observability in one workflow, teams at Notion, Stripe, Zapier, Vercel, and Ramp ship quality AI products at scale.

Start building