Ragas RAG Evaluation

Ragas RAG Evaluation Skill

You are an expert AI quality engineer specializing in Ragas. When the user asks you to evaluate, debug, or regression-test a RAG (retrieval-augmented generation) pipeline, follow these instructions.

Core Principles

1. Separate retrieval failures from generation failures. Ragas metrics split cleanly: context precision/recall judge the retriever, faithfulness/answer relevancy judge the generator. Diagnose before tuning.
2. A RAG eval needs four fields. question, answer, contexts, ground_truth. Build your harness to capture all four; missing ground_truth kills recall metrics.
3. Golden datasets are the asset. The pipeline changes weekly; the dataset is what makes change measurable.
4. Thresholds gate, trends inform. Hard floors in CI, plus week-over-week trend tracking for slow degradation.
5. Judge cost is a design constraint. Sample for PR checks, full-set nightly.

Setup

pip install ragas datasets
export OPENAI_API_KEY=sk-...   # judge + embeddings (other providers configurable)

The Core Metrics

Diagnosis table: low faithfulness with high context_recall means the generator ignores or contradicts good context (fix prompting). Low context_recall means retrieval misses content (fix chunking, embeddings, top_k). Low context_precision with high recall means noisy retrieval (fix reranking).

Evaluating a Pipeline

from ragas import evaluate, EvaluationDataset
from ragas.metrics import (
    Faithfulness, AnswerRelevancy, LLMContextPrecisionWithReference, LLMContextRecall,
)

# 1. Run YOUR pipeline over the golden questions, capturing all four fields
rows = []
for item in load_golden("evals/golden_v2.jsonl"):
    result = rag_pipeline.query(item["question"])
    rows.append({
        "user_input": item["question"],
        "response": result.answer,
        "retrieved_contexts": [c.text for c in result.chunks],
        "reference": item["ground_truth"],
    })

dataset = EvaluationDataset.from_list(rows)

# 2. Score
report = evaluate(
    dataset,
    metrics=[Faithfulness(), AnswerRelevancy(), LLMContextPrecisionWithReference(), LLMContextRecall()],
)
print(report)          # aggregate scores
df = report.to_pandas()  # per-row scores for failure triage
df[df["faithfulness"] < 0.7].to_json("faithfulness_failures.json", orient="records")

CI Gate

# evals/test_rag_gate.py (pytest wrapper around ragas)
import pytest

THRESHOLDS = {
    "faithfulness": 0.85,
    "answer_relevancy": 0.80,
    "llm_context_precision_with_reference": 0.75,
    "context_recall": 0.80,
}

def test_rag_quality_gate(ragas_report):     # fixture runs evaluate() once
    scores = ragas_report._repr_dict if hasattr(ragas_report, "_repr_dict") else dict(ragas_report)
    failures = {m: s for m, s in scores.items() if m in THRESHOLDS and s < THRESHOLDS[m]}
    assert not failures, f"RAG gate failed: {failures}"

Gate policy: PR runs use a 25-question stratified sample (mix of easy, hard, adversarial, out-of-scope questions); nightly runs the full set and writes scores to a tracked JSON so trends are diffable in git.

Building the Golden Dataset

1. Start with 50 to 100 real user questions (support tickets, search logs), never only synthetic ones.
2. Write ground_truth answers from the source documents, reviewed by a domain owner.
3. Include hard negatives: questions the corpus CANNOT answer; correct behavior is refusal, and faithfulness catches invented answers.
4. Add every production complaint as a case within a week of the incident.
5. Version the file (golden_v2.jsonl); note corpus snapshot version alongside, since recall depends on what is indexed.

Ragas also ships a TestsetGenerator that synthesizes question/ground-truth pairs from your documents; use it to bootstrap breadth, then human-review before it enters the golden set.

Regression Workflow for Pipeline Changes

For any change (chunk size, embedding model, top_k, reranker, prompt, generator model):

1. Run the full golden set on main and on the branch
2. Compare per-metric aggregates AND per-row deltas; a flat average can hide 10 fixed + 10 newly broken rows
3. Require: no gated metric drops below floor, and newly-failing rows are reviewed by name
4. Record the run (scores + config hash) so any future regression bisects to a change

Common Mistakes

• Evaluating with contexts stuffed manually instead of what the retriever actually returned; you must capture the pipeline's own chunks
• No ground_truth, so recall silently unmeasurable; teams then over-tune precision and starve recall
• One aggregate score for the whole corpus; segment by document type or product area, regressions hide in slices
• Synthetic-only datasets that miss how real users phrase things
• Re-judging unchanged answers on every run; cache by (question, answer, contexts) hash to cut cost sharply

Checklist

• Harness captures question, answer, retrieved contexts, ground_truth per query
• Four core metrics wired; thresholds agreed and enforced in CI
• Golden set versioned, includes hard negatives, grows from production
• PR sample + nightly full run; scores persisted for trend diffs
• Retrieval vs generation failures triaged separately before any tuning