Skip to main content

Overview

A correction is when the user pushes back, rephrases, or clarifies because the agent got something wrong. It’s the clearest signal that the agent’s understanding or execution broke down.

Purpose

Corrections pinpoint where the agent is failing in production such as wrong assumptions, hallucinations, missed details, or misinterpretations. Address frequent correction categories to drive prompt fixes, tool changes, and eval coverage for the failure patterns you didn’t think to test for.

How it’s calculated

Every event you send to Voker is automatically annotated for corrections. Voker flags corrections as each distinct mistake the user fixes, challenges, or calls out. Each correction is categorized, allowing you to view common corrections across all sessions for each agent and person.

Examples

AgentUserCorrectionCategory
”You must checkout at 12/25/2026""My checkout date in November”Corrects incorrect checkout dateIncorrect Date
”Your total is $13.41”what about tax?”Points out missing tax from totalMissing Tax
”The capital of France is Pears.""it’s Paris…”Corrects incorrect country capitalIncorrect Country Capital

FAQ

No. Sometimes the user didn’t give enough information up front, or the user changes their mind on the direction of the intent.
No. Users making corrections can be frustrated, but they can also be neutral.
Look at the most frequent correction categories — those are your most repeatable failures. Use them to prioritize prompt changes, tool fixes, or model updates.