Overview
Automated annotations are Voker’s analytics layer for AI agents in production. Every event you send to Voker is automatically labeled with what the user was trying to accomplish, where the agent got something wrong, and whether the user’s intent was fulfilled.
Annotation types
Voker produces three kinds of automated annotations.Intents
What the user is trying to accomplish.
Corrections
Where the agent got something wrong.
Resolutions
Whether an intent was actually fulfilled.
Categorization
We automatically categorize these annotations into broader Categories. Read more about this process in the Categorizations documentation.FAQ
Do automated annotations improve over time?
Do automated annotations improve over time?
Definitely! We are constantly improving our automated annotations system so that it’s immediately helpful out of the box.
Are automated annotations customizable?
Are automated annotations customizable?
Yes. Contact the team to customize how intents, corrections, and resolutions are detected to reflect your product’s reality instead of a generic default.
How frequently are automated annotations updated?
How frequently are automated annotations updated?
Every event you send is annotated automatically.
How can I export my automated annotations?
How can I export my automated annotations?
Contact the team to get your exported annotations.
Why do I need automated annotations?
Why do I need automated annotations?
Logs and traces help engineers debug one conversation at a time. Automated annotations tell you what’s happening across all of them — what users are asking, where the agent keeps breaking, and whether agents actually deliver. They close the measurement gap that existing observability and product analytics tools weren’t built for.
Do automated annotations replace evals or manual annotations?
Do automated annotations replace evals or manual annotations?
No. Evals catch regressions against scenarios you’ve already defined; automated annotations surface failure patterns in production that you didn’t know to write evals for. They are complementary.