Data Drift is the change over time in the statistical properties of inputs to a deployed AI system, eroding model accuracy even when the underlying task is unch
Data drift is change in the input distribution a system sees versus what it was trained or tuned on: new vocabulary, shifting user demographics, seasonal patterns, upstream schema changes. Detection compares live feature/embedding distributions against baselines with statistical distance metrics, alerting when divergence crosses thresholds.
Drift is the slow leak of production AI: models stay green on dashboards while accuracy decays on inputs they no longer recognize. Monitoring inputs, not just outputs, gives early warning before users feel the failure, and tells you when retraining or re-indexing is due.
Data drift: inputs change, relationships hold (new question mix, same correct answers). Concept drift: the input→output relationship itself changes (same transactions, new fraud meaning). Different diagnoses, different fixes.
Embed incoming queries and compare distributions to baselines, track topic mixes, watch retrieval hit rates, and sample outputs through evals. Rising 'no good retrieval' rates are a classic early drift signal.
Diagnose source first: refresh retrieval corpora, update few-shot examples or prompts, retrain/re-tune on recent data, or fix upstream pipeline changes. Blind retraining without diagnosis wastes cycles and can chase noise.