Interpret RAG drift scores, recommend thresholds, and explain 5 drift dimensions via @mukundakatta/ragdrift-mcp — 3 tools

@mukundakatta/ragdrift-mcp v0.1.1 — verified recipe

Install & run: npm install @mukundakatta/ragdrift-mcp then spawn node <prefix>/node_modules/@mukundakatta/ragdrift-mcp/src/index.js via stdio.

3 tools

5 dimensions

• data — per-feature distribution shift (KS + PSI) on tabular features (latency, retrieval count, token count)
• embedding — distribution shift in embedding space (MMD² with RBF kernel + Sliced Wasserstein-1)
• response — length and semantic shift in model responses (KS on lengths + Sliced Wasserstein on embeddings)
• confidence — confidence distribution + calibration shift (KS + ECE delta)
• query — workload composition shift (k-means clustering + symmetric KL divergence)

Severity scale

Key gotchas

1. This is a pure reference/diagnostic server — it does NOT detect drift from raw data. You provide the drift score (from your own detector), it interprets and recommends.
2. `recommend_thresholds` scales by sqrt(1000/n) — larger samples get tighter thresholds. sample_size minimum is 50.
3. `false_positive_budget` range is 0.005–0.5, default 0.05. It adjusts strictness multiplicatively.
4. All responses are structured JSON — easy to parse programmatically.
5. `threshold` param in `interpret_drift_score` adds `exceeded: true/false` to the response.
6. Sub-millisecond after JIT: p50 = 0.4ms, first call ~1.3ms.

Verified trace (11 calls, 100% success)

explain_drift_dimensions({})                                    → 5-dimension reference (1.3ms)
interpret_drift_score({score:0.02, dimension:"embedding"})      → "no significant shift" (0.6ms)
interpret_drift_score({score:0.85, dimension:"data"})           → "severe shift, action required" (0.4ms)
interpret_drift_score({score:0.35, dimension:"response", threshold:0.3}) → "significant shift", exceeded:true (0.2ms)
interpret_drift_score({score:0.15, dimension:"confidence"})     → "moderate shift, watch closely" (0.5ms)
interpret_drift_score({score:0.5, dimension:"query"})           → "severe shift, action required" (0.3ms)
recommend_thresholds({dimension:"embedding"})                   → {conservative:0.25, moderate:0.5, lax:1} (0.5ms)
recommend_thresholds({dimension:"data", sample_size:10000, false_positive_budget:0.01}) → {conservative:0.0375, moderate:0.075, lax:0.15} (0.5ms)
recommend_thresholds({dimension:"confidence", sample_size:100, false_positive_budget:0.1}) → {conservative:0.21, moderate:0.42, lax:0.84} (0.4ms)
interpret_drift_score({score:0, dimension:"query"})             → "no significant shift" (0.2ms)
interpret_drift_score({score:1.0, dimension:"data"})            → "severe shift, action required" (0.2ms)