@wavequery/conductor/dist/cjs/evaluation/eval-runner.js

Modular LLM orchestration framework

"use strict";
Object.defineProperty(exports, "__esModule", { value: true });
exports.EvalRunner = void 0;
const logger_1 = require("@/utils/logger");
const server_1 = require("@/visualization/server");
const evaluators_1 = require("./evaluators");
class EvalRunner {
    constructor(config) {
        this.config = config;
        this.logger = new logger_1.Logger({
            level: logger_1.LogLevel.DEBUG,
            prefix: "EvalRunner",
        });
        this.evaluators = [
            new evaluators_1.AutomatedEvaluator(config),
            new evaluators_1.HumanEvaluator(config),
            new evaluators_1.CustomEvaluator(config),
        ];
        if (config.visualization) {
            this.initializeVizServer();
        }
    }
    initializeVizServer() {
        this.vizServer = new server_1.VizServer(undefined, {
            theme: "light",
        });
    }
    async runEvaluation(input, expected) {
        const startTime = Date.now();
        const allResults = [];
        try {
            for (const evaluator of this.evaluators) {
                try {
                    const results = await evaluator.evaluate(input, expected);
                    allResults.push(...results);
                    if (this.config.visualization) {
                        await this.visualizeResults(results);
                    }
                }
                catch (error) {
                    this.logger.error(`Evaluator failed:`, error);
                }
            }
            const aggregatedResults = this.aggregateResults(allResults);
            this.logger.info("Evaluation completed", {
                duration: Date.now() - startTime,
                totalMetrics: aggregatedResults.length,
            });
            return aggregatedResults;
        }
        catch (error) {
            this.logger.error("Evaluation failed:", error);
            throw error;
        }
    }
    async visualizeResults(results) {
        if (!this.vizServer)
            return;
        const nodes = results.map((result) => ({
            id: `metric-${result.metricName}`,
            type: 'metric',
            label: result.metricName,
            data: {
                score: result.score,
                metadata: result.metadata,
            },
            status: result.score >= (this.config.threshold || 0.7) ? "completed" : "warning",
        }));
        this.vizServer.updateGraph({ nodes, edges: [] });
    }
    aggregateResults(results) {
        const { aggregation = "mean" } = this.config;
        if (aggregation === "weighted") {
            return results.map((result) => {
                const weight = this.config.metrics.find((m) => m.name === result.metricName)
                    ?.weight || 1;
                return {
                    ...result,
                    score: result.score * weight,
                };
            });
        }
        return results;
    }
    async cleanup() {
        if (this.vizServer) {
            await this.vizServer.close();
        }
    }
}
exports.EvalRunner = EvalRunner;
//# sourceMappingURL=eval-runner.js.map