@dooor-ai/toolkit/dist/evals/latency.js

Guards, Evals & Observability for AI applications - works seamlessly with LangChain/LangGraph

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.LatencyEval = void 0;
const base_1 = require("./base");
/**
 * Evaluates the latency of LLM responses
 *
 * Score calculation:
 * - latency <= target: score = 1.0
 * - target < latency < max: linear scale from 1.0 to 0.0
 * - latency >= max: score = 0.0
 */
class LatencyEval extends base_1.Eval {
    constructor(config = {}) {
        super(config);
        this.thresholdMs = config.thresholdMs ?? 3000; // 3 seconds
        this.targetMs = config.targetMs ?? 1000; // 1 second
        this.maxMs = config.maxMs ?? 10000; // 10 seconds
    }
    get name() {
        return "LatencyEval";
    }
    evaluate(input, output, metadata) {
        const latency = metadata?.latency ?? 0;
        // Calculate score
        let score;
        if (latency <= this.targetMs) {
            score = 1.0;
        }
        else if (latency >= this.maxMs) {
            score = 0.0;
        }
        else {
            // Linear interpolation between target and max
            const range = this.maxMs - this.targetMs;
            const delta = latency - this.targetMs;
            score = 1.0 - (delta / range);
        }
        const passed = latency <= this.thresholdMs;
        return {
            name: this.name,
            score: Math.max(0, Math.min(1, score)), // Clamp to [0, 1]
            passed,
            details: passed
                ? `Latency ${latency}ms is acceptable (threshold: ${this.thresholdMs}ms)`
                : `Latency ${latency}ms exceeds threshold of ${this.thresholdMs}ms`,
            metadata: {
                latency,
                thresholdMs: this.thresholdMs,
                targetMs: this.targetMs,
                maxMs: this.maxMs,
            },
            timestamp: new Date(),
        };
    }
}
exports.LatencyEval = LatencyEval;