ragdrift-core

Pure-Rust core for ragdrift: 5-dimensional drift detection for production RAG systems.

This crate is the numerical engine. It has no Python or FFI surface and is suitable for embedding in a Rust service that wants to compute drift scores without going through Python. The Python bindings live in ragdrift-py in the same workspace.

What it computes

• Data drift — feature-wise Kolmogorov-Smirnov and Population Stability Index.
• Embedding drift — Maximum Mean Discrepancy (RBF kernel, median-heuristic bandwidth) plus sliced Wasserstein-1 over the embedding matrix.
• Response drift — distribution shift over response length and unigram entropy, with optional embedding-based semantic shift.
• Confidence drift — KS test on score distributions plus Expected Calibration Error delta.
• Query-pattern drift — k-means cluster assignment shift (KL divergence).

Quick example

use ndarray::Array2;
use ragdrift_core::detectors::EmbeddingDriftDetector;

let baseline: Array2<f32> = Array2::zeros((1024, 384));
let current: Array2<f32> = Array2::zeros((1024, 384));
let detector = EmbeddingDriftDetector::default();
let score = detector.detect(&baseline.view(), &current.view()).unwrap();
println!("{} = {:.4}", score.method, score.score);

License

Dual-licensed under MIT or Apache-2.0 at your option.