DriftWatch JavaScript SDK

npm install thread-driftwatch

Semantic drift and model staleness monitor for AI systems. Part of the Thread Suite.

Quick Start

const DriftWatch = require("thread-driftwatch");
const dw = new DriftWatch(); // defaults to https://drift-watch.onrender.com

// Register your model endpoint
const endpoint = await dw.createEndpoint("my-model", "https://your-model.com/run");

// Define a ground truth anchor
const anchor = await dw.createAnchor(
  "tax-law-ghana",
  "What is the current VAT rate in Ghana?",
  "15%",
  "https://gra.gov.gh"
);

// Run a drift check
const result = await dw.runCheck(endpoint.id, "tax-law-ghana");

console.log(result.staleness_score);  // 1.0 = fully current, 0.0 = fully stale
console.log(result.drift_detected);   // true = a previously passing anchor is now failing

Methods

Fact Anchors

dw.createAnchor(domain, question, expectedContains, sourceUrl, verifiedAt, contributedBy)
dw.listAnchors(domain)
dw.getAnchor(anchorId)
dw.updateAnchor(anchorId, data)
dw.deactivateAnchor(anchorId)

Model Endpoints

dw.createEndpoint(name, url, owner)
dw.listEndpoints()
dw.getEndpoint(endpointId)
dw.deleteEndpoint(endpointId)

Drift Checks

dw.runCheck(endpointId, domain)   // run all anchors against live endpoint
dw.listChecks(endpointId)         // check history
dw.getCheck(checkId)              // full check with per-anchor results
dw.getStaleness(endpointId, domain) // latest staleness score + trend

Utils

dw.stats()   // dashboard overview
dw.health()  // health check

What It Detects

DriftWatch detects semantic drift — when an AI model that used to answer a question correctly starts getting it wrong. This happens silently when:

• Model providers push updates without notice
• Laws, regulations, or facts change after training cutoff
• Company policies or pricing change and the model doesn't know

Links

• Live API
• API Docs
• GitHub
• Python SDK
• Thread Suite