RAG Eval

Evaluate RAG systems by diagnosing retrieval and generation separately. A single "RAG accuracy" number hides whether the problem is finding the right documents or using them correctly. This skill separates them so you know what to fix.

When to Activate

• User has a RAG pipeline (retriever + generator) with traces
• User wants to know if their RAG system hallucinates
• User is optimizing chunking strategy and needs before/after comparison
• User wants to build a RAG evaluation dataset

Checklist

You MUST create a task for each item and complete them in order:

1. Load RAG traces — validate query + context + answer triples
2. Separate retrieval vs generation — determine which layers to evaluate
3. Run faithfulness evaluation — is the answer grounded in retrieved docs?
4. Run retrieval evaluation — are the right documents retrieved?
5. Build diagnostic matrix — cross-tabulate to find root cause
6. Output findings — prioritized issues with concrete fixes

Fast path: run the bundled script

Once each trace has a faithfulness judgment (and ideally a retrieval signal), build the retrieval-vs-generation diagnostic matrix with the bundled, tested script (scripts/rag_diagnostic.py, standard library only, no OpenJudge dependency):

python scripts/rag_diagnostic.py --traces traces.jsonl

Trace rows: {"faithful":bool} or {"faithfulness_score":1-5} (>= 4 = faithful), plus an optional retrieval signal {"retrieval_good":bool} or {"recall_at_k":0-1} (>= 0.5 = good). It prints the generation faithful/hallucinating split, the 2×2 matrix when a retrieval signal is present, and the primary issue (retrieval vs generation). --self-test to verify it.

Steps below explain how to separate the layers and produce the faithfulness/retrieval signals (with OpenJudge graders or any judge).

Step 1: Load RAG Traces

The minimum data needed per trace:

# Each trace must contain:
trace = {
    "query": "What is the return policy?",
    # retrieved_docs are dicts with id + text (the id is required for retrieval
    # metrics like Recall@k; the text is required for the faithfulness check).
    "retrieved_docs": [
        {"id": "doc_1", "text": "Returns are accepted within 30 days..."},
        {"id": "doc_2", "text": "Refunds are issued to the original..."},
    ],
    "answer": "You can return items within 30 days for a full refund.",
    "reference_answer": "Our 30-day return policy allows full refunds.",  # optional
    "gold_doc_ids": ["doc_1", "doc_3"],  # optional — which docs should have been retrieved
}

If your traces only have raw strings (no ids), wrap them first so retrieval metrics work: retrieved_docs = [{"id": f"d{i}", "text": t} for i, t in enumerate(raw_strings)].

Validate data completeness and report any gaps. If > 20% of traces are missing key fields, ask the user to confirm the schema before proceeding.

Step 2: Separate Retrieval vs Generation

RAG failures come from two independent sources:

Why separate them? A system with perfect retrieval but poor generation needs prompt engineering. A system with poor retrieval needs chunking/embedding work. Treating them as one problem wastes effort.

Step 3: Faithfulness Evaluation

Use OpenJudge HallucinationGrader to check if the answer stays grounded in retrieved documents:

from openjudge.graders.common.hallucination import HallucinationGrader
from openjudge.runner.grading_runner import GradingRunner

faithfulness_grader = HallucinationGrader(model=model)

runner = GradingRunner(
    grader_configs={"faithfulness": faithfulness_grader},
    max_concurrency=8,
)

# Dataset format for HallucinationGrader
dataset = [
    {
        "query": trace["query"],
        "response": trace["answer"],
        "context": "\n\n".join(doc["text"] for doc in trace["retrieved_docs"]),
    }
    for trace in traces
]

results = await runner.arun(dataset)

# HallucinationGrader scores 1-5 (5 = no hallucination, fully grounded)
# Binarize: score >= 4 → faithful, score < 4 → hallucination

Also evaluate answer relevance — does the response actually address the query?

from openjudge.graders.common.relevance import RelevanceGrader

relevance_grader = RelevanceGrader(model=model)

Step 4: Retrieval Evaluation

If gold_doc_ids are available, compute retrieval metrics:

def recall_at_k(retrieved_docs, gold_ids, k=5):
    """Fraction of gold docs found in top-k retrieved docs."""
    retrieved_ids = set(doc["id"] for doc in retrieved_docs[:k])
    gold_set = set(gold_ids)
    if not gold_set:
        return None
    return len(retrieved_ids & gold_set) / len(gold_set)

def precision_at_k(retrieved_docs, gold_ids, k=5):
    """Fraction of top-k docs that are relevant."""
    retrieved_ids = set(doc["id"] for doc in retrieved_docs[:k])
    gold_set = set(gold_ids)
    if not retrieved_ids:
        return 0
    return len(retrieved_ids & gold_set) / len(retrieved_ids)

def mrr(retrieved_docs, gold_ids):
    """Mean Reciprocal Rank — how early the first relevant doc appears."""
    for i, doc in enumerate(retrieved_docs):
        if doc["id"] in gold_ids:
            return 1 / (i + 1)
    return 0

If gold_doc_ids are NOT available, use a semantic relevance judge as a proxy: check whether each retrieved doc is semantically related to the query.

Chunking Optimization

If retrieval is the bottleneck, try a grid search over chunk size and overlap:

chunk_configs = [
    {"chunk_size": 256, "overlap": 32},
    {"chunk_size": 512, "overlap": 64},
    {"chunk_size": 1024, "overlap": 128},
    {"chunk_size": 512, "overlap": 128},
]

for config in chunk_configs:
    # Re-chunk, re-embed, re-retrieve, compute Recall@5
    ...

Step 5: Diagnostic Matrix

The most valuable output — cross-tabulate retrieval vs generation results:

                    Generation: Faithful    Generation: Hallucinating
Retrieval: Good          ████████ 62%            ██ 10%
                         (system works)          (generator problem)

Retrieval: Poor          █ 8%                   ██████ 20%
                         (lucky guess)           (both broken)

Interpretation:

• 62% (Good/Faithful): System working correctly
• 10% (Good/Hallucinating): Generator is misusing or fabricating despite having the right docs. Fix: prompt engineering, few-shot examples, or model upgrade.
• 8% (Poor/Faithful): Generator got the right answer from wrong docs — likely using parametric knowledge, not retrieval. Unreliable.
• 20% (Poor/Hallucinating): Both layers broken. Fix retrieval first, then generation.

Step 6: Output

Present findings with concrete recommendations:

RAG Evaluation Results (500 traces):

Layer Results:
  Faithfulness:     76% pass  (95% CI: [72%, 80%])
  Answer Relevance: 82% pass  (95% CI: [78%, 85%])
  Retrieval Recall@5: 70%    (95% CI: [66%, 74%])

Diagnostic Matrix:
                    Faithful    Hallucinating
  Retrieval Good      62%          10%  ← 10% of answers hallucinate despite good docs
  Retrieval Poor       8%          20%  ← 20% both broken

Primary issue: 10% hallucination rate even with good retrieval.
  → Generator is ignoring or misreading retrieved documents.
  → Recommendation: add "cite specific document passages" to generation prompt.
  → Re-evaluate faithfulness after prompt change.

Secondary issue: 20% of traces have both poor retrieval and hallucination.
  → Recommendation: optimize chunking first (try smaller chunks with more overlap).
  → Re-run retrieval eval after chunking changes.

Common Mistakes

• Using a single "RAG accuracy" metric. This is the cardinal sin of RAG eval. You can't fix what you can't localize. Always separate retrieval from generation.
• Skipping faithfulness because correctness looks OK. An answer can be correct (matches reference) but unfaithful (not derived from retrieved docs). This means the generator is using parametric knowledge — which will fail silently on a different knowledge domain.
• Optimizing generation before retrieval. If retrieval is broken, no amount of prompt engineering will fix the answers. Fix the data pipeline first.
• No adversarial questions in the eval set. RAG systems look great on factual lookups but fail on queries that require synthesizing across multiple chunks. Include multi-hop and comparative questions.

Next Skills

After 05-rag-eval:

• 03-align-human: Calibrate the faithfulness judge against human judgments.
• 01-eval-design: Build a more comprehensive RAG eval dataset with adversarial queries.
• 06-prompt-regression: If you change the generation prompt, measure before/after.