LLM Evaluation — Measuring Output Quality

# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
# LLM EVALUATION METRICS & APPROACHES
# ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

# ── 1. ROUGE (for summarization) ──────────────────────
from rouge_score import rouge_scorer

scorer = rouge_scorer.RougeScorer(["rouge1", "rouge2", "rougeL"], use_stemmer=True)
reference = "The cat sat on the mat and fell asleep."
generated = "The cat lay down on the mat and went to sleep."
scores = scorer.score(reference, generated)
for metric, score in scores.items():
    print(f"  {metric:8s}: P={score.precision:.3f} R={score.recall:.3f} F1={score.fmeasure:.3f}")

# ROUGE-1: unigram overlap  (word-level)
# ROUGE-2: bigram overlap   (phrase-level)
# ROUGE-L: longest common subsequence (sentence structure)

# ── 2. LLM-AS-JUDGE — GPT-4 evaluates GPT-4-mini output ──
from openai import OpenAI
client = OpenAI()

def llm_judge(question: str, answer: str, criteria: list[str]) -> dict:
    """Use GPT-4o to evaluate quality of an answer."""
    criteria_str = "\n".join(f"{i+1}. {c}" for i, c in enumerate(criteria))

    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "user",
            "content": f"""Evaluate this answer on a scale of 1-10 for each criterion.

Question: {question}
Answer: {answer}

Criteria:
{criteria_str}

Respond as JSON: {{"scores": {{}}, "overall": 0, "reasoning": ""}}"""
        }],
        response_format={"type": "json_object"},
        temperature=0,
    )
    import json
    return json.loads(response.choices[0].message.content)

# ── 3. RAGAS — Evaluation for RAG pipelines ─────────────
# pip install ragas
from ragas import evaluate
from ragas.metrics import (
    faithfulness,          # is the answer grounded in the retrieved context?
    answer_relevancy,      # is the answer relevant to the question?
    context_precision,     # did we retrieve the right documents?
    context_recall,        # did we retrieve ALL relevant documents?
)

# RAGAS eval dataset format:
eval_data = {
    "question":   ["What is the return policy?"],
    "answer":     ["Items can be returned within 30 days with a receipt."],
    "contexts":   [["Our policy: items can be returned within 30 days of purchase with original receipt."]],
    "ground_truth": ["30-day return policy with receipt required."]
}

# from datasets import Dataset
# ragas_dataset = Dataset.from_dict(eval_data)
# results = evaluate(ragas_dataset, metrics=[faithfulness, answer_relevancy, context_precision, context_recall])
# print(results)  # scores for each metric

# ── 4. HALLUCINATION DETECTION ─────────────────────────
def check_hallucination(question: str, answer: str, source_docs: list[str]) -> bool:
    """Use LLM to detect if the answer contains information not in sources."""
    context = "\n".join(source_docs)
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "user",
            "content": f"""Does this answer contain claims NOT supported by the context?
Context: {context}
Answer: {answer}
Respond only: HALLUCINATION or GROUNDED"""
        }],
        temperature=0, max_tokens=10,
    )
    return "HALLUCINATION" in response.choices[0].message.content.upper()