TKYO Drift is a lightweight, transparent drift tracking library designed to monitor AI workflows, detecting changes in text-based AI interactions over time to help identify when models or users evolve their behavior.
Features
AI Drift Detection
Tracks the semantic and conceptual drift in AI models by comparing embeddings over time to a configurable baseline, detecting shifts in AI response to inputs.
Scalar Metrics Tracking
Calculates metrics like punctuation density and entropy to detect shifts in text structure and tone, allowing users to observe non-semantic changes in AI model outputs.
One-off and Training Ingestion
Allows input of individual and batch data for drift analysis, supporting both rolling and fixed baseline comparisons.
Production-Safe Integration
Runs asynchronously to track drift in backend environments without impacting production performance by logging drift details in the background.