RAG Evaluation Metrics Cheat Sheet

RAG metrics broadly fall into four layers:

Can the Retriever find it?
Is the Context complete and clean?
Does the Generator answer correctly and stay grounded in evidence?
Is the whole system safe, fast, and economical?

Traditional retrieval metrics are usually calculated directly from manually labeled relevant documents and do not require an LLM Judge. NVIDIA’s current Retriever evaluation includes Precision, Recall, nDCG, MAP, Reciprocal Rank, and Success@k. Ragas, AWS, and Databricks add metrics for complete RAG systems, including Context, Groundedness, Correctness, and Citation.

@k means that only the first k retrieval results are checked. For example, Recall@10 looks only at the first 10 results.


1. Retrieval and Ranking Metrics

MetricWhy it mattersRough calculation
Hit@k / Success@kChecks whether at least one correct piece of evidence appears in the first k results; useful for single-evidence questionsRecord 1 when Top-k contains at least one relevant result and 0 otherwise, then average across all questions
Recall@kChecks whether required evidence was missed; it is one of the most important baseline metrics for a RAG RetrieverNumber of relevant pieces of evidence found in Top-k ÷ total number of relevant pieces that should be found
Precision@kChecks how many retrieved results are genuinely useful and reflects the amount of noiseNumber of relevant results in Top-k ÷ k
MRRChecks how early the first correct result appears; useful when finding one authoritative answer is enoughTake the reciprocal of the rank of the first relevant result; rank 2 is 1/2, then average across all questions
MAP@kConsiders both whether multiple relevant results are found and whether they are ranked highlyCompute Precision whenever a relevant result appears, then average those Precision values across all questions
nDCG@kDistinguishes core evidence, supporting evidence, and partially relevant evidence rather than treating all relevance as binary; especially useful for evaluating a RerankerLabel results with different relevance grades; more important results ranked higher contribute more, then compare with the ideal ranking
All-Evidence@kChecks whether the complete evidence chain required by a multi-hop question was foundRecord 1 when every required evidence group enters Top-k; missing any group gives 0
Document Recall@kChecks whether the correct documents were found and helps locate whether a problem is at the document or Chunk levelNumber of correct documents hit in Top-k ÷ number of correct documents that should be hit
Evidence/Chunk Recall@kChecks whether the passages inside the correct documents that actually support the answer were foundGolden Evidence Chunks or evidence ranges hit in Top-k ÷ all required evidence

Quick Intuition

Hit@k:
Was at least one result found?

Recall@k:
How much of the evidence that should be found was found?

Precision@k:
How much of what was retrieved is genuinely useful?

nDCG@k:
Was the most important evidence ranked first?

All-Evidence@k:
Was the complete evidence chain found?

All-Evidence@k is a useful custom engineering metric, but unlike Recall, nDCG, and MAP, it does not have one universally standardized name. It is best designed around labeled “evidence groups,” allowing multiple equivalent pieces of evidence to satisfy the same requirement.


2. Context Quality Metrics

Traditional Retriever metrics generally depend on manually labeled relevant documents. Context metrics more often use a reference answer or an LLM Judge to determine whether the Context finally passed to the generation model is clean, complete, and sufficient.

MetricWhy it mattersRough calculation
Context PrecisionChecks whether the final Context contains too much noise and whether relevant Chunks are ranked firstJudge each Chunk’s usefulness; the more relevant Chunks there are and the earlier they appear, the higher the score
Context RecallChecks whether all facts required by the reference answer can be found in the ContextSplit the reference answer into factual claims and calculate how many are supported by the retrieved Context
Context CoverageSimilar to Context Recall, but emphasizes whether the Context covers the information required by the reference answerInformation from the reference answer covered by Context ÷ all information required by the reference answer
Context RelevanceChecks whether retrieved content is genuinely related to the user’s questionUsually ask an LLM Judge to score the relevance of each Chunk to the question, then average the scores
Context SufficiencyChecks whether the current Context is already sufficient to produce a complete answerUse expected_facts or a reference answer to judge whether the Context supports a complete response, usually with binary or graded scoring
Context Noise RatioChecks how many tokens are occupied by irrelevant content, which affects quality, latency, and costNumber of irrelevant Context tokens ÷ total Context tokens
Duplicate Context RateChecks whether the same information is repeatedly passed to the model, avoiding duplicate bias and wasted tokensNumber of duplicate or highly similar Chunks ÷ total number of final Context Chunks

Context Recall vs. Recall@k

Recall@k:
Calculated from manually labeled relevant documents or evidence IDs.

Context Recall:
Based on facts in the reference answer, judging whether those facts can be supported by the Context.

3. Answer Generation Quality Metrics

MetricWhy it mattersRough calculation
Answer CorrectnessChecks whether the final answer is correct in its facts and conclusions; it is the most direct outcome metricCompare the answer with a reference answer or expected_facts and calculate the proportion of correct facts, or have an LLM Judge score it against a Rubric
Faithfulness / GroundednessChecks whether the answer stays faithful to the retrieved evidence and helps detect hallucinationsSplit the answer into factual claims and calculate how many are supported by the Context
CompletenessPrevents an answer from being correct but covering only part of the questionRequired facts in the reference answer that are covered by the actual answer
Answer Relevance / Response RelevancyChecks whether the answer genuinely responds to the user’s question instead of being correct but off-topicJudge the relevance of the answer to the user’s intent and penalize irrelevant or repetitive content
HelpfulnessChecks whether the answer can genuinely help the user complete a taskUsually scored by a human or LLM Judge based on usefulness, clarity, and actionability
Logical CoherenceChecks for contradictions, logical jumps, or inconsistencies inside the answerInspect the answer against a Rubric for complete and consistent reasoning
Exact MatchSuitable for fixed-form answers such as model names, dates, numbers, and enum valuesRecord 1 when the normalized system answer exactly matches the reference answer and 0 otherwise
Fact-level AccuracyMore suitable for enterprise RAG than whole-paragraph text similaritySplit the answer into multiple facts and calculate correct facts ÷ total evaluated facts

Correctness and Faithfulness Are Different

Correctness:
Is the answer correct in reality or according to the reference answer?

Faithfulness:
Can the answer be supported by the retrieved Context?

For example:

The answer is correct, but the Context does not support it.
→ The model may have guessed correctly from its own memory.

The answer is faithful to the Context, but the answer is wrong.
→ The knowledge-base document may be outdated or incorrect.

4. Citation Quality Metrics

MetricWhy it mattersRough calculation
Citation PrecisionPrevents citations from existing but failing to support the corresponding claimNumber of citations that correctly support the corresponding claim ÷ total citations
Citation Coverage / Citation RecallPrevents a system from giving a few correct citations while leaving most facts uncitedFactual claims with valid citations ÷ all claims that require citations
Citation ValidityChecks whether citations genuinely exist and whether links and page numbers are validNumber of valid citations ÷ total citations
Citation EntailmentChecks whether the cited source actually entails the conclusion in the answerClassify each claim-citation pair as supported, unsupported, or contradicted

For example:

The answer contains 10 facts.
Only 2 are cited, and both citations are correct.

Citation Precision = high
Citation Coverage = low

5. Refusal and Answerability Metrics

It is not enough to evaluate whether an answer is good; the system must also be evaluated on whether it knows when to refuse.

MetricWhy it mattersRough calculation
Refusal AccuracyChecks whether the system refuses in the right situationsCorrect answers or correct refusals ÷ all cases
Unsupported Answer RateMeasures how often the system answers anyway when the documents provide no evidenceCases where the system answers despite an unanswerable question ÷ total unanswerable questions
Over-refusal RateMeasures how often the system refuses even though sufficient evidence existsCases incorrectly refused among answerable questions ÷ total answerable questions
Clarification AccuracyChecks whether the system correctly asks for clarification when a question is ambiguousCases that should trigger clarification and actually do so ÷ all cases that should trigger clarification

The ideal state is:

Evidence exists → Answer correctly
No evidence → Refuse
Ambiguity exists → Clarify first
No permission → Deny access

6. Security, Authorization, and Production Metrics

These are not necessarily standardized academic metrics, but they should usually be hard Release Gates for enterprise deployment.

MetricWhy it mattersRough calculation
Unauthorized Retrieval RateChecks whether the Retriever returned Chunks the user is not authorized to accessUnauthorized results ÷ total retrieved results
Cross-tenant Leakage RatePrevents tenant A from receiving tenant B’s dataTests with cross-tenant exposure ÷ total authorization tests
Document Injection Success RateChecks whether malicious document instructions can control the modelCases successfully manipulated by document injection ÷ total injection tests
p50 / p95 / p99 LatencyThe average can hide a small number of very slow requests; production usually focuses more on p95Sort request latencies; p95 means 95% of requests do not exceed that latency
Cost per QueryPrevents a small quality improvement from causing a large increase in token and model costsTotal Retriever, Reranker, Embedding, and LLM cost ÷ request count
Token per QueryChecks whether Context is too long or contains duplication and wasteAverage or percentile of input and output tokens used per request
Error RateMonitors timeouts, model errors, and parsing failuresFailed requests ÷ total requests

Security metrics should normally require:

Unauthorized Retrieval Rate = 0
Cross-tenant Leakage Rate = 0

Permission leakage cannot be offset by improving other quality metrics.


When first building a RAG system, there is no need to implement every metric at once. Start with this set:

Evaluation layerPriority metric
Did retrieval miss anything?Recall@k
Is ranking reasonable?nDCG@k
Was the complete evidence found?All-Evidence@k
Is the Context clean?Context Precision
Is the Context sufficient?Context Recall or Context Sufficiency
Is the answer correct?Answer Correctness
Is it grounded in evidence?Faithfulness / Groundedness
Was information omitted?Completeness
Are citations reliable?Citation Precision + Citation Coverage
Does the system make things up without evidence?Unsupported Answer Rate
Does it refuse too often?Over-refusal Rate
Production performancep95 Latency + Cost per Query
Authorization safetyUnauthorized Retrieval Rate

This combination can be remembered as:

Recall:
Did we find everything?

nDCG:
Is the ranking good?

All-Evidence:
Is the evidence chain complete?

Context Precision:
How much noise is there?

Correctness:
Is the answer correct?

Faithfulness:
Is the answer based on evidence?

Completeness:
Did the answer omit anything?

Citation:
Are citations correct and complete?

Refusal:
Will the system make things up without evidence?

Latency / Cost / ACL:
Can the system safely enter production?

References