@nestjs-agentic/evaluation

Evaluation, benchmarking, position-debiasing, and trajectory inspection primitives for autonomous NestJS AI agents.

Overview

The @nestjs-agentic/evaluation package provides rigorous statistical benchmarking and quality evaluation tools for AI agents built with nestjs-agentic.

Core Capabilities

• Pairwise Position-Debiased LLM-as-a-Judge (PairwiseDebiasedJudge):
  • Implements the position-swap debiasing protocol from UC Berkeley LMSYS (Zheng et al., NeurIPS 2023 — MT-Bench).
  • Evaluates candidate agent responses in forward $(A, B)$ and reverse $(B, A)$ positions to detect and eliminate systematic primacy/recency bias.
• Trajectory Efficiency & Tool Precision Metrics (TrajectoryInspectorMetric, ToolPrecisionMetric):
  • Implements trajectory evaluation protocols from Tsinghua University (Liu et al., ICLR 2024 — AgentBench).
  • Step Efficiency: Computes ratio of minimal optimal steps to executed steps ($E_{\text{step}} = N_{\text{optimal}} / N_{\text{actual}}$).
  • Tool Precision: Computes ratio of error-free tool invocations ($P_{\text{tool}} = N_{\text{successful}} / N_{\text{total}}$).
• Safety Policy Compliance (SafetyPolicyMetric):
  • Validates forbidden tools, RBAC/ABAC policy adherence, and unauthorized invocation attempts.
• Accuracy Ground Truth (AccuracyGroundTruthMetric):
  • Vector cosine similarity and Sørensen-Dice token overlap metrics.
• Benchmark Suite Runner (BenchmarkRunner) & Reporting (EvalReporter):
  • Multi-trial variance analysis ($\mu$, $\sigma$, pass-rate) with automated Markdown/JSON benchmark reports.

Installation

npm install @nestjs-agentic/evaluation nestjs-agentic

Quick Start

1. Pairwise Position-Debiased Evaluation (MT-Bench)

import { runPairwiseDebiasedJudge } from '@nestjs-agentic/evaluation';

const result = await runPairwiseDebiasedJudge(
  {
    query: 'How to handle idempotency in banking transactions?',
    candidateA: {
      id: 'agent_cascade_frugal',
      output: 'Use distributed unique idempotency keys with atomic DB locking.',
    },
    candidateB: {
      id: 'agent_legacy',
      output: 'Just check if the transfer exists in memory.',
    },
    criteria: 'Technical accuracy, robustness, and architectural soundness.',
  },
  async (query, first, second, criteria) => {
    // Invoke your preferred LLM judge model (e.g. OpenAI / Anthropic)
    const judgeVerdict = await modelAdapter.generate({
      system: 'You are an impartial judge evaluating two technical answers.',
      prompt: `Task: ${query}\nCandidate 1: ${first.output}\nCandidate 2: ${second.output}`,
    });
    return parseJudgeVerdict(judgeVerdict);
  },
);

console.log('Winner:', result.winner); // 'candidate_a'
console.log('Debiased Score A:', result.debiasedScoreA);
console.log('Position Bias Detected:', result.positionBiasDetected);

2. Trajectory Efficiency & Tool Precision Inspection (AgentBench)

import {
  ToolPrecisionMetric,
  TrajectoryInspectorMetric,
} from '@nestjs-agentic/evaluation';

const inspector = new TrajectoryInspectorMetric({ penalizeExtraSteps: true });
const precisionMetric = new ToolPrecisionMetric({ minPrecisionThreshold: 0.8 });

const evalItem = {
  id: 'transfer_task',
  query: 'Transfer $500 to account ACC-2',
  optimalSteps: 2,
  expectedToolSequence: ['checkBalance', 'transferFunds'],
  expectedToolArgs: {
    transferFunds: { amount: 500 },
  },
};

const inspectorResult = inspector.evaluate(evalItem, agentResult);
const precisionResult = precisionMetric.evaluate(evalItem, agentResult);

console.log('Trajectory Passed:', inspectorResult.passed);
console.log('Step Efficiency:', inspectorResult.details.stepEfficiency);
console.log('Tool Precision:', precisionResult.details.precision);

Academic References

1. MT-Bench / LMSYS: Zheng et al., "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena" (UC Berkeley LMSYS, NeurIPS 2023, arXiv:2306.05685)
2. AgentBench: Liu et al., "AgentBench: Evaluating Large Language Models as Agents" (Tsinghua University, ICLR 2024, arXiv:2308.03688)

License

MIT © irzix