jyjeanne/ai-model-evaluation

Purpose

To systematically assess the performance, accuracy, and safety of LLM outputs using quantitative metrics and "LLM-as-a-Judge" patterns, ensuring production readiness.

When to Use

• Before deploying any LLM application to production.
• When comparing different models (e.g., GPT-4o vs. Claude 3.5 Sonnet) or prompt versions.
• To detect regressions after updating prompts or RAG knowledge bases.

Procedure

1. Define the Evaluation Dataset (Golden Set)

Create a tests.json file containing inputs and expected outputs.

[
  {
    "input": "What is the return policy?",
    "expected": "You can return items within 30 days.",
    "context": "Our policy allows returns for 30 days from purchase date."
  }
]

2. Implementation with Promptfoo (CLI)

Promptfoo is a popular tool for running batch evaluations.

# Install
npm install -g promptfoo

# Initialize
promptfoo init

Configure promptfooconfig.yaml:

prompts:
  - "Answer this question using the context: {{context}}. Question: {{input}}"

providers:
  - openai:gpt-4o

tests:
  - vars:
      input: "What is the return policy?"
      context: "30-day return policy applies."
    assert:
      - type: icontains
        value: "30 days"
      - type: llm-rubric
        value: "Does not mention unrelated topics"

3. RAG Evaluation (Ragas/DeepEval)

For RAG systems, evaluate the three-way relationship: Question, Context, and Answer.

import { rce } from "deepeval"; // Conceptual example

async function evaluateRag(query: string, retrievalContext: string, output: string) {
  // 1. Faithfulness: Is the answer grounded in the context?
  // 2. Answer Relevance: Does it answer the query?
  // 3. Context Precision: Was the retrieved context relevant?
}

4. Custom LLM-as-a-Judge Script

Use a stronger model to grade your target model.

async function gradeOutput(question: string, answer: string, reference: string) {
  const graderPrompt = `
    You are an impartial judge. Grade the student's answer based on the reference.
    Question: ${question}
    Reference: ${reference}
    Student Answer: ${answer}
    
    Provide a score from 1-10 and a brief explanation.
    Output JSON: { "score": number, "explanation": string }
  `;
  
  // Call GPT-4 with JSON mode enabled
}

5. Continuous Integration (CI)

Integrate evaluation into your GitHub Actions to prevent regressions.

# .github/workflows/ai-eval.yml
jobs:
  evaluate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npx promptfoo eval

Constraints

• Bias: LLM judges tend to prefer longer answers or answers from the same provider. Use diverse judges (OpenAI + Anthropic).
• Cost: Running evaluations on 1000s of rows can be expensive. Use gpt-4o-mini for simpler checks.
• Reference Accuracy: A "Golden Set" is only as good as the human-verified reference answers.

Expected Output

A detailed report (HTML/JSON) showing pass/fail status, accuracy percentages, and regression analysis.