Definition: Model drift is the gradual decline in a deployed model's performance because the data it sees, or the relationship between inputs and outcomes, has changed since it was built or last evaluated.
Data drift happens when inputs change, for example new products, customer segments or document formats. Concept drift happens when the meaning of the data changes, such as fraud patterns that adapt to detection. For LLM applications, behavior can also shift when a hosted model version, prompt or retrieval source changes.
Teams track input distributions, prediction confidence and, where labels arrive later, real accuracy. For language models they rerun fixed evaluation sets on a schedule and after every model or prompt update, and sample production traces for review. Alerts fire when metrics move past agreed thresholds.
Responses include retraining on recent data, updating prompts or retrieval content, rolling back a model version, or adding human review while a fix is tested. A sound data pipeline makes retraining routine rather than an emergency.
Ryz MLOps engineers and machine learning engineers build drift monitoring and retraining pipelines, working on your team.
Thanks — your message has been sent. We’ll get back to you soon.
Something went wrong while sending your message. Please try again or email info@ryzlabs.com.