19 — RAG Evaluation Fundamentals¶

Learn how to systematically evaluate Retrieval-Augmented Generation (RAG) systems across retrieval quality, context quality, answer quality, grounding, citations, latency, cost, and overall user experience.

📖 Overview¶

Building a RAG pipeline is only the beginning.

A system can successfully execute:

Query
  ↓
Retrieval
  ↓
Context
  ↓
LLM
  ↓
Answer

and still produce poor results.

For example:

User Question
      ↓
Retriever
      ↓
Wrong Documents
      ↓
LLM
      ↓
Confident Wrong Answer

Therefore, production RAG systems require systematic evaluation.

The central question is not:

"Did the pipeline run successfully?"

It is:

"Did the system retrieve the right evidence and use that evidence to produce a correct, relevant, grounded, and useful answer?"

1. Why RAG Evaluation Matters¶

A RAG system contains multiple stages:

Query
 ↓
Query Processing
 ↓
Embedding
 ↓
Retrieval
 ↓
Filtering
 ↓
Ranking
 ↓
Context Assembly
 ↓
Prompt
 ↓
LLM
 ↓
Answer

A failure at any stage can affect the final response.

For example:

Poor Chunking
      ↓
Poor Retrieval
      ↓
Poor Context
      ↓
Poor Answer

Therefore, evaluating only the final answer is insufficient.

2. RAG Evaluation Dimensions¶

A practical evaluation model is:

                 RAG Evaluation

        ┌─────────────────────────┐
        │     Retrieval Quality   │
        └────────────┬────────────┘
                     ↓
        ┌─────────────────────────┐
        │      Context Quality    │
        └────────────┬────────────┘
                     ↓
        ┌─────────────────────────┐
        │      Generation Quality │
        └────────────┬────────────┘
                     ↓
        ┌─────────────────────────┐
        │     System Quality      │
        └─────────────────────────┘

The major dimensions are:

Retrieval
Context
Generation
Grounding
Citations
Latency
Cost
Reliability
User Experience

3. Retrieval vs Generation Evaluation¶

One of the most important distinctions is:

Did we retrieve the correct information?

versus:

Did the LLM correctly use the retrieved information?

Example:

Question:
"What is the annual leave entitlement?"

Retrieved:
"Employees receive 25 days."

Generated:
"Employees receive 30 days."

Retrieval succeeded.

Generation failed.

4. Retrieval Failure Example¶

Question:
"What is the annual leave entitlement?"

Retrieved:
"Employees receive 10 sick days."

Generated:
"Employees receive 10 sick days."

The LLM may have correctly summarized the supplied context.

But the RAG system still failed.

The problem is:

Retrieval Quality

5. RAG Evaluation Pipeline¶

flowchart TD
    A["Evaluation Dataset"] --> B["RAG Pipeline"]
    B --> C["Retrieved Context"]
    B --> D["Generated Answer"]

    A --> E["Expected Evidence"]
    A --> F["Reference Answer"]

    C --> G["Retrieval Evaluation"]
    C --> H["Context Evaluation"]

    D --> I["Generation Evaluation"]
    D --> J["Grounding Evaluation"]
    D --> K["Citation Evaluation"]

    G --> L["Evaluation Report"]
    H --> L
    I --> L
    J --> L
    K --> L

    B --> M["Latency / Cost Metrics"]
    M --> L

6. Evaluation Dataset¶

A RAG evaluation requires representative questions.

Example:

[
  {
    "question": "What is the annual leave entitlement?",
    "expected_source": "employee-handbook",
    "expected_section": "Annual Leave",
    "reference_answer": "Employees receive 25 days of annual leave."
  },
  {
    "question": "How should annual leave be requested?",
    "expected_source": "employee-handbook",
    "expected_section": "Annual Leave",
    "reference_answer": "Employees should submit requests through the employee portal."
  }
]

The dataset becomes the foundation for repeatable evaluation.

7. What an Evaluation Dataset Should Contain¶

A useful evaluation record may include:

Question
Expected Answer
Expected Source
Expected Section
Expected Evidence
Metadata Filters
Difficulty
Question Type

Example:

{
  "id": "rag-001",
  "question": "What is the annual leave entitlement?",
  "expected_answer": "25 days",
  "expected_source": "employee-handbook",
  "expected_section": "Annual Leave",
  "category": "policy"
}

8. Evaluation Dataset Types¶

A production dataset should contain different question types.

Fact Questions
Policy Questions
Multi-Sentence Questions
Comparison Questions
Procedural Questions
Ambiguous Questions
Unanswerable Questions

For example:

"What is the annual leave entitlement?"

"How do I request annual leave?"

"Can unused leave be carried forward?"

"What is the policy for contractors?"

"Does the handbook mention remote work?"

9. Answerable vs Unanswerable Questions¶

A strong evaluation dataset should include questions for which the knowledge base contains no answer.

Example:

Question:
"What is the company's stock price?"

Knowledge Base:
Employee Handbook

Expected behavior:

No supporting evidence
        ↓
Controlled response

The system should not fabricate an answer.

10. Retrieval Evaluation¶

Retrieval evaluation asks:

Did the retriever return the relevant evidence?

Important metrics include:

Precision@K
Recall@K
MRR
NDCG
Hit Rate

These metrics measure different aspects of retrieval quality.

11. Precision@K¶

Precision@K measures how many of the top K retrieved results are relevant.

Conceptually:

Precision@K =
Relevant Results in Top-K
-------------------------
Total Results in Top-K

For example:

Top 5 Results

Relevant:
3

Precision@5:

3 / 5 = 0.60

A higher value means the retrieved result set contains more relevant documents.

12. Precision@K Formula¶

Example:

K = 5
Relevant = 4

Precision@5 = 4 / 5 = 0.80

Precision is especially useful when irrelevant context is expensive.

13. Recall@K¶

Recall@K measures whether the relevant evidence was retrieved within the top K results.

Conceptually:

Recall@K =
Relevant Results Retrieved in Top-K
-----------------------------------
Total Relevant Results

Example:

Total relevant chunks = 4
Retrieved relevant chunks = 3

Recall@K = 3 / 4 = 0.75

14. Recall@K Formula¶

Recall is important because:

If the correct evidence is never retrieved, the generation stage cannot use it.

15. Precision vs Recall¶

Precision
    ↓
"How much of what I retrieved was relevant?"

Recall
    ↓
"How much of the relevant information did I retrieve?"

Example:

Retriever A:

5 results
4 relevant

High Precision


Retriever B:

10 results
4 relevant

Lower Precision
Higher opportunity for Recall

The correct balance depends on the application.

16. Hit Rate¶

Hit Rate measures whether at least one relevant result appears within the top K.

Example:

Expected Evidence:
Chunk 42

Top 5:
Chunk 7
Chunk 12
Chunk 42
Chunk 81
Chunk 90

The query is a hit.

Hit Rate = 1

If no relevant result appears:

Hit Rate = 0

17. Mean Reciprocal Rank¶

MRR considers the position of the first relevant result.

If the first relevant document appears at:

Position 1

the reciprocal rank is:

1 / 1 = 1.0

If it appears at:

Position 4

then:

1 / 4 = 0.25

18. MRR Formula¶

MRR is useful when the position of the first relevant result matters.

19. NDCG¶

NDCG stands for:

Normalized Discounted Cumulative Gain

It considers both:

Relevance

and:

Ranking Position

A highly relevant result near the top contributes more than the same result appearing much later.

NDCG is useful when relevance can have multiple levels.

For example:

Highly Relevant
Relevant
Partially Relevant
Irrelevant

20. Retrieval Metrics Summary¶

Metric	Main Question
Precision@K	How many retrieved results are relevant?
Recall@K	Did we retrieve the relevant evidence?
Hit Rate	Did at least one relevant result appear?
MRR	How early is the first relevant result?
NDCG	Are highly relevant results ranked near the top?

No single metric completely describes retrieval quality.

21. Context Evaluation¶

Retrieval can succeed while context assembly fails.

Example:

Retriever
   ↓
Correct 5 chunks
   ↓
Context Builder
   ↓
Only 2 chunks preserved

Therefore context should also be evaluated.

Important dimensions include:

Context Relevance
Context Completeness
Context Precision
Context Ordering
Context Noise

22. Context Relevance¶

Context relevance asks:

Does the retrieved context actually help answer the question?

Example:

Question:
"What is the annual leave entitlement?"

Context:
"Employees receive 25 days of annual leave."

High relevance.

But:

Context:
"Employees should use the company cafeteria."

Low relevance.

23. Context Completeness¶

Context completeness asks:

Does the retrieved context contain enough information to answer the question?

Example:

Question:
"How many leave days can be carried forward?"

Context:
"Employees may carry forward unused leave."

If the number of days is missing, the context may be incomplete.

24. Context Noise¶

Too much irrelevant context can reduce generation quality.

Relevant Chunk
Relevant Chunk
Relevant Chunk
Irrelevant Chunk
Irrelevant Chunk
Irrelevant Chunk

The goal is not:

Maximum Context

but:

Maximum Useful Context

25. Context Quality Pipeline¶

flowchart LR
    A["Retrieved Chunks"] --> B["Relevance"]
    B --> C["Completeness"]
    C --> D["Deduplication"]
    D --> E["Token Budget"]
    E --> F["Final Context"]

26. Generation Evaluation¶

Generation evaluation asks:

Did the LLM produce a useful answer from the supplied evidence?

Important dimensions include:

Correctness
Relevance
Faithfulness
Groundedness
Completeness
Clarity
Citation Accuracy

27. Answer Correctness¶

Answer correctness asks:

Does the generated answer correctly answer the question?

Example:

Question:
"What is the annual leave entitlement?"

Expected:
25 days

Generated:
Employees receive 25 days of annual leave.

Result:
Correct

28. Answer Relevance¶

An answer can be factually correct but still poorly focused.

Question:

"What is the annual leave entitlement?"

Answer:

Employees receive 25 days of annual leave.
The company also has policies for sick leave,
remote work, travel expenses, and office security...

The answer begins correctly but contains unnecessary information.

Answer relevance measures whether the response directly addresses the user's question.

29. Faithfulness¶

Faithfulness asks:

Is the generated answer supported by the retrieved context?

Example:

Context:
Employees receive 25 days of annual leave.

Answer:
Employees receive 25 days of annual leave.

Faithful

But:

Context:
Employees receive 25 days of annual leave.

Answer:
Employees receive 30 days of annual leave.

Not faithful

30. Groundedness¶

Groundedness is closely related to faithfulness.

A grounded answer should be supported by the evidence provided to the LLM.

Retrieved Evidence
       ↓
Generated Answer
       ↓
Supported?

If the answer contains unsupported claims:

Grounding ↓

31. Hallucination in RAG¶

RAG reduces the need for the model to rely entirely on pretrained knowledge, but it does not eliminate hallucinations.

Example:

Context:
Employees receive 25 days of annual leave.

Answer:
Employees receive 25 days and can
automatically carry forward 15 days.

If the context does not mention the carry-forward limit, that claim is unsupported.

32. Groundedness Evaluation¶

A practical evaluation process can be:

Generated Answer
       ↓
Split into Claims
       ↓
Check Each Claim
       ↓
Supported by Context?
       ↓
Groundedness Score

Example:

Claim 1 → Supported
Claim 2 → Supported
Claim 3 → Unsupported

33. Claim-Level Evaluation¶

Suppose the answer is:

Employees receive 25 days of annual leave.
Requests must be submitted through the portal.
Unused leave may be carried forward.

Break it into:

Claim 1
Claim 2
Claim 3

Then evaluate each claim independently.

Claim 1 → Supported
Claim 2 → Supported
Claim 3 → Supported

This gives a more detailed view than simply marking the entire answer correct or incorrect.

34. Citation Evaluation¶

Enterprise RAG systems often return citations.

Example:

Employees receive 25 days of annual leave.

Source:
Employee Handbook — Page 42

Citation evaluation should verify:

Citation Exists
Citation Points to Correct Source
Citation Supports the Claim
Citation Is Not Fabricated

35. Citation Accuracy¶

Bad example:

Answer:
Employees receive 25 days.

Citation:
Employee Handbook — Page 90

If page 90 contains unrelated content:

Citation Accuracy = Failure

The answer may still be correct, but the source attribution is wrong.

36. Citation Coverage¶

If an answer contains several factual claims:

Claim 1 → Citation
Claim 2 → Citation
Claim 3 → No Citation

citation coverage is incomplete.

A production system should determine which claims require source attribution.

37. Answer Completeness¶

An answer can be correct but incomplete.

Question:

"How do I request annual leave?"

Context:

Employees should submit requests
through the employee portal.
Requests require manager approval.

Answer:

Use the employee portal.

Correct, but incomplete.

A better answer:

Submit the request through the employee portal.
The request then requires manager approval.

38. Answer Quality Model¶

                 Answer Quality

             ┌────────────────┐
             │   Correctness  │
             └───────┬────────┘
                     ↓
             ┌────────────────┐
             │   Relevance    │
             └───────┬────────┘
                     ↓
             ┌────────────────┐
             │  Groundedness  │
             └───────┬────────┘
                     ↓
             ┌────────────────┐
             │ Completeness   │
             └───────┬────────┘
                     ↓
             ┌────────────────┐
             │ Citation       │
             └────────────────┘

A strong answer should perform well across these dimensions.

39. Human Evaluation¶

Human evaluation remains important.

A reviewer can score an answer on:

Correctness
Relevance
Grounding
Completeness
Clarity
Citation Quality

For example:

Score: 1 → Poor
Score: 2 → Weak
Score: 3 → Acceptable
Score: 4 → Good
Score: 5 → Excellent

40. Human Evaluation Rubric¶

Dimension	Question
Correctness	Is the answer factually correct?
Relevance	Does it answer the question directly?
Groundedness	Is it supported by retrieved evidence?
Completeness	Does it include important information?
Clarity	Is it easy to understand?
Citation Quality	Are sources accurate and useful?

41. LLM-as-a-Judge¶

An LLM can also evaluate generated responses.

Conceptually:

Question
+
Retrieved Context
+
Generated Answer
        ↓
Evaluation LLM
        ↓
Score + Explanation

Example evaluation prompt:

Evaluate whether the answer is supported
by the provided context.

Question:
{question}

Context:
{context}

Answer:
{answer}

Return:

{
  "grounded": true,
  "score": 0.95,
  "reason": "The answer is directly supported by the context."
}

LLM-based evaluation should itself be validated.

42. Human Evaluation vs Automated Evaluation¶

Human Evaluation¶

Advantages:

High-quality judgment
Nuanced reasoning
Better for complex questions

Limitations:

Expensive
Slow
Subjective

Automated Evaluation¶

Advantages:

Fast
Repeatable
Scalable

Limitations:

Evaluator bias
Prompt sensitivity
Model limitations

A production evaluation strategy often combines both.

43. Reference-Based Evaluation¶

If a reference answer exists:

Question
   ↓
Expected Answer
   ↓
Generated Answer
   ↓
Comparison

This can evaluate:

Correctness
Similarity
Completeness

However, exact string matching is usually insufficient for natural-language answers.

44. Semantic Answer Evaluation¶

Two answers may express the same meaning differently.

Reference:

Employees receive 25 days of annual leave.

Generated:

The annual leave entitlement is 25 days per year.

String comparison would show a difference.

Semantic evaluation can recognize that the answers convey the same information.

45. Exact Match¶

Exact Match is useful when the expected output has a deterministic form.

Example:

Question:
"What is the policy year?"

Expected:
2026

Generated:

Exact match succeeds.

It is less suitable for open-ended answers.

46. Token-Level Metrics¶

Traditional NLP metrics can compare generated text against references.

Examples include:

BLEU
ROUGE

However, they should not be treated as complete measures of RAG quality.

A response can use different words while still being correct.

47. RAG-Specific Evaluation¶

RAG evaluation should focus on:

Retrieval
Context
Grounding
Answer
Citations

rather than relying solely on text-overlap metrics.

48. Evaluation Frameworks¶

RAG evaluation can be implemented manually or using specialized tools.

Examples include:

Ragas
DeepEval
TruLens
Custom Evaluation Pipelines

Frameworks can help automate metrics and evaluation workflows.

The important principle is:

Understand the metric before relying on a framework to calculate it.

49. Evaluation Without a Framework¶

A simple custom evaluator can be implemented.

def evaluate_answer(
    expected_answer,
    generated_answer
):

    return {
        "contains_expected": (
            expected_answer.lower()
            in generated_answer.lower()
        )
    }

This is intentionally simple.

Production evaluation should use richer semantic and grounding checks.

50. Retrieval Evaluation Example¶

def evaluate_retrieval(
    retrieved_ids,
    expected_ids,
    k=5
):

    retrieved = set(
        retrieved_ids[:k]
    )

    expected = set(
        expected_ids
    )

    relevant = (
        retrieved.intersection(expected)
    )

    precision = (
        len(relevant) / k
        if k
        else 0
    )

    recall = (
        len(relevant) / len(expected)
        if expected
        else 0
    )

    return {
        "precision_at_k": precision,
        "recall_at_k": recall
    }

This demonstrates the basic idea behind retrieval evaluation.

51. Evaluation Record¶

A useful evaluation result might look like:

{
  "question_id": "rag-001",
  "retrieval": {
    "precision_at_5": 0.8,
    "recall_at_5": 1.0,
    "hit": true
  },
  "generation": {
    "correct": true,
    "grounded": true,
    "relevant": true
  },
  "citation": {
    "accurate": true
  },
  "latency_ms": 910
}

This makes results machine-readable.

52. Batch Evaluation¶

Instead of evaluating one question:

Question 1
Question 2
Question 3
...
Question 1000

run the complete evaluation dataset.

results = []

for item in evaluation_dataset:

    result = rag_pipeline.answer(
        item["question"]
    )

    evaluation = evaluate(
        item,
        result
    )

    results.append(
        evaluation
    )

This allows regression testing.

53. Evaluation Dashboard¶

A production team can track:

Retrieval Recall@K
Retrieval Precision@K
Hit Rate
Answer Correctness
Groundedness
Citation Accuracy
Latency
Cost
Failure Rate

Example:

Retrieval Recall@5       0.91
Retrieval Precision@5    0.78
Groundedness             0.94
Answer Correctness       0.89
Citation Accuracy        0.97
P95 Latency              1.8s

These numbers are illustrative.

54. Evaluation Trend¶

Evaluation should be continuous.

flowchart LR
    A["Baseline"] --> B["Change"]
    B --> C["Run Evaluation"]
    C --> D["Compare Metrics"]
    D --> E["Accept / Reject"]
    E --> F["Production"]
    F --> G["Monitor"]
    G --> B

This turns RAG evaluation into an engineering feedback loop.

55. Regression Testing¶

Suppose the current system has:

Recall@5 = 0.91

After changing the chunking strategy:

Recall@5 = 0.73

The change should be investigated before deployment.

Similarly:

Groundedness:
0.94 → 0.81

could indicate that the new prompt or context strategy is causing problems.

56. Evaluation Before and After Changes¶

Common changes requiring evaluation include:

Embedding Model
Chunk Size
Chunk Overlap
Metadata Filters
Top-K
Similarity Threshold
Prompt
LLM Model
Context Formatting
Vector Database
Retrieval Strategy

Every change can affect downstream quality.

57. Evaluation Matrix¶

Change	Retrieval Impact	Generation Impact
Embedding Model	High	Indirect
Chunk Size	High	High
Chunk Overlap	Medium	Medium
Top-K	High	High
Similarity Threshold	High	High
Prompt	Low	High
LLM Model	None	High
Context Formatting	Low	High
Vector Database	Potentially High	Indirect

58. Evaluation by Question Type¶

Different questions may produce different results.

Example:

Fact Questions
    ↓
High Accuracy

Multi-hop Questions
    ↓
Lower Accuracy

Ambiguous Questions
    ↓
Variable Accuracy

Unanswerable Questions
    ↓
Hallucination Risk

Therefore aggregate scores alone can hide important weaknesses.

59. Segment Evaluation¶

Evaluate by categories.

Policy Questions
    Accuracy = 95%

Procedural Questions
    Accuracy = 89%

Unanswerable Questions
    Safe Response = 96%

Multi-document Questions
    Accuracy = 78%

This reveals where improvements are needed.

60. Difficulty Levels¶

Evaluation questions can also be classified:

Easy
Medium
Hard

Example:

Easy¶

"What is the annual leave entitlement?"

Medium¶

"How is annual leave requested and approved?"

Hard¶

"What happens to unused leave when an employee changes departments?"

Difficulty-based evaluation helps identify system limitations.

61. Negative Testing¶

RAG systems should be tested with intentionally difficult or unsupported queries.

Examples:

Unknown Topic
Wrong Department
Outdated Policy
Unauthorized Information
Ambiguous Query
Contradictory Documents

Expected behavior should be defined beforehand.

62. Security Evaluation¶

Security should be evaluated separately.

Example:

User:
Employee A

Query:
"What is the executive compensation policy?"

If the user does not have access:

Expected:
No unauthorized information

The test should verify that restricted content never reaches the generation layer.

63. Evaluation of Empty Retrieval¶

Test:

Question outside knowledge base

Expected:

Controlled response

Not:

Confident hallucination

This is an important production safety test.

64. Evaluation of Citation Integrity¶

Test:

Generated Claim
      ↓
Citation
      ↓
Source
      ↓
Does Source Support Claim?

If not:

Citation Failure

Citation integrity is particularly important for:

Legal
Finance
Compliance
Healthcare
Enterprise Policy

applications.

65. Latency Evaluation¶

RAG latency should be measured end-to-end.

Total Latency =
Query Processing
+
Embedding
+
Retrieval
+
Context Processing
+
LLM
+
Validation

Measure:

Average
P50
P95
P99

rather than relying only on average latency.

66. Why P95 Matters¶

Suppose:

Average latency = 800 ms
P95 latency     = 2.5 s

Most users may receive fast responses, but a significant minority experience much slower requests.

Production systems should therefore monitor latency distributions.

67. Cost Evaluation¶

A RAG request can incur:

Embedding Cost
+
Vector Infrastructure Cost
+
Reranking Cost
+
LLM Input Cost
+
LLM Output Cost

Track cost per:

Request
User
Tenant
Day
Month
Application

where useful.

68. Cost vs Quality¶

Reducing cost aggressively can reduce answer quality.

Example:

Top-K: 10
    ↓
High Context
    ↓
Higher Cost
    ↓
Potentially Better Recall

versus:

Top-K: 3
    ↓
Lower Context
    ↓
Lower Cost
    ↓
Potentially Lower Recall

The goal is not minimum cost.

It is:

Best Quality / Cost Trade-Off

69. Evaluation Scorecard¶

A practical scorecard can be:

Retrieval
  Recall@5
  Precision@5
  MRR

Context
  Relevance
  Completeness

Generation
  Correctness
  Groundedness
  Relevance

Citations
  Accuracy
  Coverage

System
  P95 Latency
  Cost / Query
  Error Rate

70. Example Scorecard¶

Category	Metric	Example
Retrieval	Recall@5	0.91
Retrieval	Precision@5	0.82
Retrieval	MRR	0.88
Context	Relevance	0.93
Generation	Correctness	0.90
Generation	Groundedness	0.95
Citations	Accuracy	0.97
System	P95 Latency	1.8s

The values are illustrative.

71. Weighted Evaluation¶

Some applications may combine metrics.

For example:

Overall Score =
0.30 × Retrieval
+
0.30 × Groundedness
+
0.20 × Correctness
+
0.10 × Citation
+
0.10 × Relevance

The weights should reflect business priorities.

For high-risk applications, correctness and grounding may deserve greater importance.

72. Why a Single Score Is Dangerous¶

A single score can hide failures.

Example:

Overall Score = 0.91

but:

Retrieval Recall = 0.98
Groundedness     = 0.95
Citation Accuracy = 0.60

The system may still be unsuitable for an enterprise application requiring reliable citations.

Always inspect individual metrics.

73. Evaluation Architecture¶

flowchart TD
    A["Evaluation Dataset"] --> B["RAG System"]

    B --> C["Retrieval Results"]
    B --> D["Generated Answers"]
    B --> E["Telemetry"]

    C --> F["Retrieval Metrics"]
    C --> G["Context Metrics"]

    D --> H["Correctness"]
    D --> I["Groundedness"]
    D --> J["Relevance"]
    D --> K["Citation Accuracy"]

    E --> L["Latency"]
    E --> M["Cost"]
    E --> N["Reliability"]

    F --> O["Evaluation Report"]
    G --> O
    H --> O
    I --> O
    J --> O
    K --> O
    L --> O
    M --> O
    N --> O

74. Production Evaluation Workflow¶

1. Define evaluation goals.

2. Build representative questions.

3. Identify expected evidence.

4. Define reference answers where appropriate.

5. Run the RAG system.

6. Capture retrieved documents.

7. Measure retrieval quality.

8. Evaluate context quality.

9. Evaluate generated answers.

10. Measure groundedness.

11. Validate citations.

12. Measure latency.

13. Measure cost.

14. Segment results by question type.

15. Compare against baseline.

16. Investigate regressions.

17. Approve or reject changes.

18. Continuously monitor production behavior.

75. Baseline Evaluation¶

Before optimizing the system, establish a baseline.

Example:

Embedding Model: Model A
Chunk Size: 500
Overlap: 50
Top-K: 5
LLM: Model X

Baseline:

Recall@5        = 0.88
Groundedness    = 0.91
Correctness     = 0.86
Citation        = 0.94
P95 Latency     = 1.9s

Future changes should be compared against this baseline.

76. Experiment Tracking¶

A RAG experiment should record:

Embedding Model
Chunk Size
Chunk Overlap
Top-K
Similarity Threshold
Vector Store
Prompt Version
LLM Model
Evaluation Dataset Version
Evaluation Metrics

Example:

{
  "experiment": "rag-v12",
  "embedding_model": "embedding-v3",
  "chunk_size": 500,
  "chunk_overlap": 50,
  "top_k": 5,
  "llm": "enterprise-model-x",
  "prompt_version": "prompt-v4"
}

This enables reproducible experimentation.

77. Evaluation Dataset Versioning¶

The dataset itself should be versioned.

eval-v1
eval-v2
eval-v3

If questions change, metric comparisons may no longer be directly comparable.

Therefore:

System Version
+
Evaluation Dataset Version

should be recorded together.

78. Production Monitoring vs Offline Evaluation¶

These are complementary.

Offline Evaluation¶

Controlled Dataset
      ↓
Repeatable Testing

Production Monitoring¶

Real User Traffic
      ↓
Real-World Behavior

Offline evaluation provides controlled comparison.

Production monitoring reveals actual system behavior.

79. Continuous Evaluation¶

flowchart LR
    A["Code Change"] --> B["Offline Evaluation"]
    B --> C{"Pass?"}

    C -->|No| D["Reject Change"]
    C -->|Yes| E["Deploy"]

    E --> F["Production Monitoring"]
    F --> G["New Evaluation Data"]
    G --> B

This creates a continuous quality loop.

80. Evaluation Alerts¶

Production systems can alert when:

Groundedness drops
Retrieval recall drops
Empty retrieval increases
Latency increases
Cost increases
Citation failures increase
Error rate increases

Example:

Groundedness:
0.94 → 0.81

This should trigger investigation.

81. Root-Cause Analysis¶

When quality decreases, inspect the pipeline stage.

Quality Drop
     ↓
Retrieval?
     ↓
Context?
     ↓
Prompt?
     ↓
LLM?
     ↓
Citation?

Do not immediately change the model.

The failure may be caused by:

Chunking
Metadata
Embedding
Filtering
Top-K
Context Budget

82. Example Root-Cause Investigation¶

Suppose:

Answer Correctness:
0.91 → 0.76

Check:

Recall@5:
0.90 → 0.61

This suggests:

Retrieval degradation

rather than an LLM generation problem.

83. Another Root-Cause Example¶

Suppose:

Recall@5:
0.91 → 0.92

Groundedness:
0.95 → 0.81

Retrieval remains strong.

The problem may be:

Prompt
Context Formatting
LLM
Response Generation

This illustrates why multiple metrics are required.

84. Evaluation Anti-Patterns¶

84.1 Only Measuring Final Answers¶

This hides retrieval failures.

84.2 Using Only One Question¶

One question cannot represent an enterprise workload.

84.3 Using Only Exact String Matching¶

Equivalent answers may use different wording.

84.4 Ignoring Unanswerable Questions¶

This hides hallucination behavior.

84.5 Ignoring Citations¶

Correct answers can still have incorrect sources.

84.6 Ignoring Latency¶

A high-quality system can still be unusable if responses are too slow.

84.7 Ignoring Cost¶

An accurate system may still be economically impractical.

84.8 Changing Multiple Variables at Once¶

If you change:

Chunking
Embedding
Prompt
LLM
Top-K

at the same time, it becomes difficult to identify what caused the improvement or regression.

85. Controlled Experiments¶

Prefer:

Change One Major Variable
        ↓
Run Evaluation
        ↓
Compare

Example:

Experiment A:
Chunk size = 500

Experiment B:
Chunk size = 800

Keep other major variables stable.

86. Evaluation of Chunk Size¶

Example experiment:

Chunk Size     Recall@5

300            0.82
500            0.90
800            0.87
1200           0.79

The best chunk size is not necessarily the largest or smallest.

It should be selected based on evaluation results and workload characteristics.

87. Evaluation of Top-K¶

Example:

Top-K     Recall@K     Groundedness

3         0.78        0.96
5         0.89        0.95
10        0.94        0.90
20        0.97        0.82

Increasing K improves retrieval coverage but may introduce context noise.

This demonstrates why retrieval and generation metrics should be evaluated together.

88. Quality Trade-Off¶

Top-K ↑
   ↓
Recall ↑
   ↓
Context Size ↑
   ↓
Noise / Cost ↑
   ↓
Potential Groundedness ↓

The optimal value is application-specific.

89. Evaluation as an Engineering Discipline¶

RAG evaluation should be treated like software testing.

Code
 ↓
Unit Tests
 ↓
Integration Tests
 ↓
Evaluation Tests
 ↓
Deployment
 ↓
Monitoring

RAG quality should not depend on manual experimentation alone.

90. Test Pyramid for RAG¶

             ┌───────────────┐
             │ Human Review  │
             └───────────────┘
           ┌─────────────────────┐
           │ End-to-End Evaluation│
           └─────────────────────┘
        ┌───────────────────────────┐
        │ Retrieval / Generation    │
        └───────────────────────────┘
     ┌─────────────────────────────────┐
     │ Unit / Integration Tests        │
     └─────────────────────────────────┘

The lower levels should provide fast feedback.

The higher levels provide broader quality validation.

91. Evaluation Strategy by Stage¶

Stage	Example Evaluation
Document Processing	Extraction accuracy
Chunking	Context completeness
Embeddings	Retrieval benchmark
Retrieval	Recall@K / Precision@K
Ranking	MRR / NDCG
Context	Relevance / completeness
Generation	Correctness / relevance
Grounding	Faithfulness
Citations	Source accuracy
System	Latency / cost / reliability

92. RAG Quality Lifecycle¶

Build
 ↓
Evaluate
 ↓
Identify Weakness
 ↓
Change One Component
 ↓
Evaluate Again
 ↓
Compare
 ↓
Deploy
 ↓
Monitor
 ↓
Collect New Cases
 ↓
Expand Dataset
 ↓
Evaluate Again

This turns evaluation into a continuous improvement process.

93. Key Takeaways¶

RAG evaluation must measure more than whether the pipeline executes successfully.
Retrieval and generation should be evaluated separately.
Retrieval evaluation determines whether the right evidence was found.
Precision@K measures how many retrieved results are relevant.
Recall@K measures whether relevant evidence was retrieved.
Hit Rate measures whether at least one relevant result was retrieved.
MRR evaluates the position of the first relevant result.
NDCG evaluates both relevance and ranking position.
Context evaluation measures relevance, completeness, and noise.
Generation evaluation measures correctness, relevance, completeness, and clarity.
Faithfulness and groundedness measure whether answers are supported by retrieved evidence.
Citation evaluation verifies that sources are accurate and actually support generated claims.
Evaluation datasets should contain representative, difficult, and unanswerable questions.
Human evaluation provides nuanced judgment but is expensive.
Automated evaluation provides scale and repeatability but requires validation.
LLM-as-a-judge can help evaluate semantic qualities but should not be blindly trusted.
Exact-match metrics are useful for deterministic outputs but are insufficient for open-ended RAG answers.
Offline evaluation and production monitoring serve different purposes and should complement each other.
Evaluation datasets should be versioned.
System configurations should be tracked with evaluation results.
Baselines are essential for measuring improvements and regressions.
Changes to chunking, embeddings, retrieval, prompts, or LLMs should be evaluated before production rollout.
Aggregate scores can hide important weaknesses, so metrics should also be segmented by question type.
RAG evaluation should include latency, cost, reliability, and security behavior.
Empty retrieval and unanswerable questions are important tests for hallucination resistance.
Evaluation should become part of the RAG engineering lifecycle rather than a final one-time activity.

The central principle is:

A production RAG system should be evaluated as a complete pipeline: retrieve the right evidence, construct useful context, generate a correct and grounded answer, provide trustworthy citations, and meet the required latency, reliability, security, and cost targets.

Part IV — Prompt Engineering & RAG Fundamentals¶

Previous Chapter: 18. Building Your First RAG Pipeline

Current Chapter: 19 — RAG Evaluation Fundamentals

Next Chapter: 20. Enterprise Generative AI Application Architecture

Part IV Chapters¶

References¶

Retrieval-Augmented Generation evaluation research
Information retrieval evaluation documentation
Precision and Recall documentation
MRR and NDCG evaluation documentation
RAG evaluation methodology documentation
Ragas documentation
DeepEval documentation
TruLens documentation
LangChain evaluation documentation
LlamaIndex evaluation documentation
LLM evaluation documentation
Enterprise AI evaluation and observability documentation

Enterprise AI Engineering Handbook
Building Production-Grade Enterprise AI Systems — One Chapter at a Time.