Chapter 20- Grounding and Faithfulness

Chapter Goal

In Chapter 19, we reached the relationship:

Evidence
+
Claim
↓
Relationship Analysis
↓
Entailment
Contradiction
Neutral

We learned that a strong factual Claim should ideally be:

Evidence
↓
Entailment
↓
Claim

But an AI answer usually contains more than one Claim.

For example:

Enterprise customers can disable renewal
from Billing Settings.

Refund requests are available within 14 days.

After that period, cancellation does not
automatically create refund eligibility.

Each statement may need support.

This raises a larger question:

How do we evaluate the answer as a whole?

This chapter introduces three related but distinct concepts:

Groundedness
Faithfulness
Factuality

Groundedness

Is the answer supported
by the available Evidence?

Faithfulness

Does the answer accurately reflect
the supplied Evidence?

Factuality

Is the information actually true
in the real world?

These concepts can overlap, but they are not identical.

Consider:

Evidence:
Company X was founded in 2018.

AI Answer:
Company X was founded in 2018.

If the Evidence supports the answer:

Grounded

If the answer accurately represents the Evidence:

Faithful

If Company X really was founded in 2018:

Factually Correct

Ideally, we want all three.

But they can come apart.

For example, imagine the supplied Evidence itself is outdated and incorrectly says:

Current refund period = 30 days

The AI faithfully answers:

The current refund period is 30 days.

Relative to the supplied Evidence, the answer may be:

Grounded
Faithful

while in the real world the current policy is actually:

14 days

So the answer may still be:

Factually Incorrect

This distinction is central to evaluating RAG systems.

By the end of this chapter, you should understand:

• what Grounding means;
• what Groundedness means;
• what a Grounded Answer is;
• why Grounding is not necessarily a separate post-generation checking stage;
• what Faithfulness means;
• what Evidence Faithfulness means;
• how Faithfulness differs from Groundedness;
• what Factuality means;
• what Factual Accuracy means;
• how Factuality differs from Evidence Support;
• how an answer can be grounded but factually wrong;
• how an answer can be factually correct but ungrounded;
• what an Unsupported Claim is in the context of hallucination;
• what Source Misinterpretation is;
• what Evidence Misuse is;
• what Citation Hallucination is;
• and how these failure modes relate to RAG quality.

The central distinction of the chapter is:

Groundedness
→ Is the answer supported by available Evidence?

Faithfulness
→ Does the answer accurately reflect that Evidence?

Factuality
→ Is the answer actually true?

20.1 Grounding

1. What Is Grounding?

Grounding means keeping the generated answer tied to available external information or Evidence.

In a RAG system:

User Question
↓
Retrieval
↓
Evidence
↓
Evidence enters LLM Context
↓
Generation
↓
Answer based on Evidence

The important relationship is:

Evidence
↓
supports
↓
Generated Claims

Grounding therefore describes the use of Evidence as the informational basis for generation.

2. Grounding Is About Support

Suppose the Evidence says:

Enterprise customers may request refunds
within 14 days of purchase.

The AI answers:

Enterprise customers can request refunds
within 14 days.

The Claim is supported by the Evidence.

The answer is grounded with respect to that Claim.

Now suppose the AI says:

Enterprise customers have a 30-day refund period.

The available Evidence does not support that statement.

The Claim is not grounded in the supplied Evidence.

Grounding
→ connect Claims to Evidence

3. Grounding Is Not the Same as Retrieval

Retrieval only finds information.

Query
↓
Retrieval
↓
Evidence Candidate

Grounding concerns what happens when the answer is generated from that information.

Evidence
↓
Generation
↓
Supported Claim

Retrieval
≠
Grounding

A system can retrieve the correct Evidence and still produce an unsupported answer.

Correct Evidence:
Refund period = 14 days

Generated Answer:
Refund period = 30 days

Retrieval succeeded.

Grounding failed.

4. Grounding Is Not Automatically Guaranteed by Context Presence

Suppose the LLM Context contains:

Passage A:
Enterprise refunds are available within 14 days.

But the answer says:

Enterprise refunds are available within 30 days.

The correct Evidence was present.

But the model did not follow it.

Evidence in Context
≠
Grounded Answer

Context availability gives the model the opportunity to use the Evidence.

Grounding depends on whether generated Claims remain supported by it.

5. Grounding Is Not Necessarily a Separate Post-Generation Stage

A common mental model is:

Generate Answer
↓
Compare Answer to Evidence
↓
Ground Answer

That is not the definition of Grounding.

Grounding refers to generation remaining tied to Evidence.

Evidence
↓
Available During Generation
↓
Generated Claims Stay Supported
↓
Grounded Answer

A system may later perform:

Generated Answer
+
Evidence
↓
Check Support

That process is better described as:

Groundedness Evaluation

or:

Groundedness Checking

Grounding
→ answer stays tied to Evidence

Groundedness Evaluation
→ check whether that actually happened

6. What Is Groundedness?

Groundedness describes the degree to which Claims in an answer are supported by the available Evidence.

Suppose an answer contains:

Claim 1
Claim 2
Claim 3

and Evidence supports:

Claim 1 → supported
Claim 2 → supported
Claim 3 → unsupported

The answer is not fully grounded.

Answer
↓
Claims
↓
Compare Each Claim With Evidence
↓
Supported / Unsupported
↓
Groundedness

7. Groundedness Can Be Evaluated at Claim Level

Consider:

AI Answer:

Enterprise costs €500 per year.
It includes priority support.
It has a 30-day refund period.

Evidence supports:

€500 per year
priority support
14-day refund period

Then:

Claim 1
→ Grounded

Claim 2
→ Grounded

Claim 3
→ Not Grounded

This produces a partially grounded answer.

Groundedness therefore does not always need to be treated as a simple yes-or-no property.

8. What Is a Grounded Answer?

A Grounded Answer is an answer whose relevant factual Claims are supported by the available Evidence.

Evidence
↓
Supports Claim 1
Supports Claim 2
Supports Claim 3
↓
Grounded Answer

Ideally:

All Important Claims
↓
Supported by Evidence

The answer should not introduce unsupported details beyond what the Evidence establishes.

9. Grounded Does Not Mean Copied

A grounded answer does not need to reproduce the Evidence word for word.

Evidence:

Refund requests must be submitted
within fourteen calendar days.

Answer:

Customers have 14 days to submit
a refund request.

The wording changed.

The meaning remains supported.

Grounded
≠
Verbatim Copy

Grounding is semantic, not lexical.

10. Groundedness Depends on Claim Meaning

Evidence:

Some customers may qualify for refunds.

Answer:

All customers qualify for refunds.

The words are similar.

But the meaning changed from:

some

to:

all

So the Claim is not grounded.

Groundedness depends on semantic support, not surface similarity.

11. Groundedness and Entailment

Chapter 19 introduced:

Evidence
+
Claim
↓
Entailment
Contradiction
Neutral

Groundedness builds on this relationship.

If an important Claim is:

Entailed

then it has strong Evidence support.

If it is:

Neutral

the Evidence does not establish it.

If it is:

Contradicted

the Evidence points against it.

A well-grounded answer should therefore contain Claims that are primarily:

Entailed by Evidence

12. Groundedness Is Evidence-Relative

This is one of the most important ideas in the chapter.

Groundedness asks:

Is the answer supported by the Evidence available to the system?

It does not directly ask:

Is the Evidence itself correct about the world?

Suppose the supplied Evidence incorrectly says:

Company X was founded in 2017.

The AI answers:

Company X was founded in 2017.

Relative to that Evidence, the answer may be grounded.

But if the company was actually founded in 2018:

Grounded
but
Factually Incorrect

20.2 Faithfulness

13. What Is Faithfulness?

Faithfulness describes whether the generated answer accurately represents the information contained in the supplied Evidence.

The central question is:

Did the model preserve the meaning of the Evidence when generating the answer?

Consider:

Evidence:
Refund requests may be submitted within 14 days.

Faithful Answer:
Customers may submit refund requests
within 14 days.

The wording changed.

The meaning remained intact.

14. Faithfulness Is About Representation

Suppose the Evidence says:

Some Enterprise customers may qualify
for additional support.

The AI says:

All Enterprise customers receive
additional support.

The answer did not faithfully represent the Evidence.

It transformed:

some

into:

all

Faithfulness
→ preserve Evidence meaning

15. What Is Evidence Faithfulness?

Evidence Faithfulness refers specifically to whether the generated answer accurately reflects the Evidence supplied to the model.

Evidence
↓
LLM
↓
Generated Claim
↓
Does Claim accurately reflect Evidence?

If yes:

Faithful

If the model changes important meaning:

Unfaithful

16. Faithfulness Can Fail Through Overstatement

Evidence:

The product may improve performance
in some workloads.

Answer:

The product always improves performance.

The answer added:

always

and removed:

some workloads

This is unfaithful to the Evidence.

Another example:

Evidence:
The policy may apply to eligible customers.

Answer:
The policy applies to every customer.

Again, the generated Claim is stronger than the Evidence.

17. Faithfulness Can Fail Through Omitted Conditions

Evidence:

Refunds are available within 14 days
if less than 80% of service credits have been used.

Answer:

Refunds are available within 14 days.

The generated statement omitted an important condition.

Omitted Qualification
↓
Potential Faithfulness Failure

The Claim may sound plausible while no longer accurately representing the complete Evidence.

18. Faithfulness Can Fail Through Distortion

Evidence:

The service is unavailable in two regions.

Answer:

The service is unavailable in most regions.

The model changed the meaning.

Unfaithfulness can involve:

overstatement
understatement
missing conditions
changed quantities
changed entities
changed certainty
misinterpreted relationships

19. Faithfulness vs. Groundedness

These concepts are closely related.

Groundedness

Is the Claim supported by the available Evidence?

Faithfulness

Does the Claim accurately reflect
what the Evidence says?

For many simple factual Claims, these judgments strongly overlap.

Evidence:
Refund period = 14 days

Claim:
Refund period = 14 days

The Claim is both:

Grounded
Faithful

20. Why Keep Groundedness and Faithfulness Separate?

Because they emphasize different failure perspectives.

Groundedness
→ support

Faithfulness
→ accurate representation

Imagine several Evidence Passages describe a complex policy.

The AI uses the correct Evidence but distorts one exception during summarization.

Correct Evidence Selected
↓
Incorrect Interpretation
↓
Unfaithful Answer

The Evidence was available, but its meaning was not preserved correctly.

21. Faithfulness Is Also Evidence-Relative

Faithfulness does not directly tell us whether the Evidence itself is true.

Suppose the supplied source incorrectly says:

The company has 10 offices.

The AI says:

The company has 10 offices.

The answer accurately reflects the Evidence.

Faithful

But if the company actually has 12 offices:

Factually Incorrect

Faithfulness
≠
External Truth

20.3 Factuality

22. What Is Factuality?

Factuality concerns whether the information stated in an answer is actually true.

The central question is:

Does the Claim correspond to reality?

For example:

Claim:
The Eiffel Tower is in Paris.

If the statement corresponds to reality:

Factually Correct

If the answer says:

The Eiffel Tower is in Berlin.

it is:

Factually Incorrect

23. What Is Factual Accuracy?

Factual Accuracy describes how accurately factual statements in an answer correspond to the real-world facts they describe.

Claim
↓
Compare With Reality / Reliable Ground Truth
↓
Correct or Incorrect

This differs from comparing the Claim only with supplied Evidence.

24. Factuality vs. Groundedness

This is one of the most important distinctions in the course.

Groundedness

Is the Claim supported by
the available Evidence?

Factuality

Is the Claim actually true?

These are not always the same.

25. Grounded and Factually Correct

Evidence:

Company X was founded in 2018.

Reality:

Company X was founded in 2018.

AI:

Company X was founded in 2018.

The answer is:

Grounded
+
Factually Correct

This is the ideal case.

26. Grounded but Factually Incorrect

Suppose an outdated source says:

Current refund period = 30 days.

The current real policy is:

14 days.

The model answers:

The current refund period is 30 days.

Relative to the supplied Evidence:

Grounded

But relative to reality:

Factually Incorrect

Groundedness
does not guarantee
Factuality

27. Factually Correct but Ungrounded

Suppose the supplied Evidence says only:

Company X was founded in 2018.

The AI answers:

Company X was founded in 2018
and is headquartered in Amsterdam.

Assume both facts are actually true.

Founded in 2018
→ Grounded + Factually Correct

Headquartered in Amsterdam
→ Factually Correct
  but Unsupported by supplied Evidence

The second Claim can therefore be:

Factually Correct
but
Ungrounded

28. Factually Correct Does Not Mean Faithful

Truth and accurate representation of Evidence are separate evaluation dimensions.

Suppose Evidence says:

Only Premium customers receive benefit X.

A faithful answer is:

Premium customers receive benefit X.

But an answer such as:

All customers receive benefit X.

misrepresents the supplied Evidence.

Even if an external circumstance happened to make the statement true at some later point, the answer would still not faithfully represent the Evidence that was supplied.

Factuality
≠
Faithfulness

29. Three Independent Questions

When evaluating a Claim, we can ask:

1. Groundedness
Is it supported by the available Evidence?

2. Faithfulness
Does it accurately represent that Evidence?

3. Factuality
Is it actually true?

The ideal result is:

Grounded
+
Faithful
+
Factually Correct

But different combinations are possible.

30. A Useful Comparison Matrix

Case A

Evidence correct
Answer reflects Evidence
Reality matches Evidence

→ Grounded
→ Faithful
→ Factually Correct

Case B

Evidence wrong or outdated
Answer reflects Evidence exactly

→ Grounded
→ Faithful
→ Factually Incorrect

Case C

Evidence does not contain the fact
Answer states a true fact from elsewhere

→ Ungrounded relative to supplied Evidence
→ Not Evidence-Supported
→ Potentially Factually Correct

Case D

Evidence correct
Answer distorts Evidence

→ Ungrounded or weakly grounded
→ Unfaithful
→ Potentially Factually Incorrect

This is why evaluating AI answers requires more than one metric.

31. Factuality Requires a Truth Reference

To determine Groundedness, we compare:

Claim
↔
Available Evidence

To determine Factuality, we need some reference for reality.

That may involve:

trusted external sources
verified databases
authoritative records
ground-truth datasets
current factual verification

Groundedness Evaluation
→ compare Claim to Evidence

Factuality Evaluation
→ compare Claim to reality / ground truth

32. Factuality Can Be Difficult to Evaluate

Real-world truth is not always simple.

Sources may:

disagree
be outdated
contain errors
describe different conditions
use different definitions

For example:

How many employees does Company X have?

Different sources may report different values because of:

different dates
contractor inclusion
subsidiary definitions
rounding
reporting delays

Factuality evaluation may therefore require careful source interpretation.

20.4 Hallucination Revisited

33. What Is Hallucination in This Context?

In AI systems, hallucination broadly refers to generated information that is unsupported, incorrectly derived, or otherwise not justified by the available informational basis.

In RAG, hallucination is not limited to completely invented facts.

Failures can occur even when relevant Evidence was retrieved.

This chapter focuses on four important forms:

Unsupported Claim
Source Misinterpretation
Evidence Misuse
Citation Hallucination

34. Unsupported Claim

We introduced the Unsupported Claim in Chapter 19.

It occurs when the AI states something that is not adequately supported by available Evidence.

Example:

Evidence:
Enterprise refunds are available within 14 days.

AI:
Enterprise customers receive a guaranteed
full refund within 14 days.

Unsupported components include:

guaranteed
full

Evidence
↓
does not establish
↓
Entire Claim

35. Unsupported Claims Can Be Completely Invented

Evidence:

Refund period = 14 days.

AI:

The company also offers a free
30-day trial.

Nothing in the supplied Evidence supports that statement.

This is a straightforward Unsupported Claim.

36. Unsupported Claims Can Be Subtle

Not all hallucinations look dramatic.

Evidence:

The product is available in Amsterdam.

AI:

The product is widely available
throughout the Netherlands.

The model expanded a narrow fact into a broader Claim.

This is still unsupported.

Hallucination can arise through:

added specificity
added certainty
broader scope
missing conditions
invented numbers
invented relationships

37. What Is Source Misinterpretation?

Source Misinterpretation occurs when the model misunderstands what a Source or Passage actually says.

Example:

Source:
Customers who cancel within 14 days
may request a refund.

AI Interpretation:
Customers must wait 14 days before
requesting a refund.

The Source was available.

But its meaning was misunderstood.

Correct Source
+
Incorrect Interpretation
=
Source Misinterpretation

38. Source Misinterpretation Can Reverse Meaning

Evidence:

Cancellation does not automatically
create refund eligibility.

AI:

Cancellation automatically creates
refund eligibility.

This is not a retrieval failure.

The correct Passage was retrieved.

The failure happened during interpretation or generation.

Correct Retrieval
↓
Correct Evidence
↓
Incorrect Interpretation
↓
Incorrect Claim

39. Source Misinterpretation Can Involve Conditions

Evidence:

The discount applies only to annual plans.

AI:

The discount applies to all plans.

The model ignored:

only to annual plans

Conditions and qualifiers are common sources of interpretation failure.

40. What Is Evidence Misuse?

Evidence Misuse occurs when real Evidence is used to support a Claim that it does not appropriately justify.

Suppose:

Evidence:
Company X operates in Amsterdam.

AI:

Company X is the largest provider
in Amsterdam.

The Evidence is genuine.

But it does not support:

largest

The model used valid Evidence for an invalid inference.

That is Evidence Misuse.

41. Evidence Misuse vs. Source Misinterpretation

Source Misinterpretation

The model misunderstands
what the Evidence says.

Evidence Misuse

The model uses Evidence
to justify something it does not support.

Example:

Evidence:
Company X has an office in Amsterdam.

Misinterpretation:

Company X has no office in Amsterdam.

Misuse:

Company X is the biggest company in Amsterdam.

The first reverses the Evidence.

The second stretches the Evidence beyond what it establishes.

42. Evidence Misuse Can Involve Wrong Entity Mapping

Evidence:

Standard Plan:
Refund period = 30 days.

AI:

Enterprise Plan has a 30-day refund period.

The factual attribute came from real Evidence.

But it was attached to the wrong entity.

Correct Evidence
+
Wrong Entity Mapping
=
Evidence Misuse

43. Evidence Misuse Can Involve Wrong Time or Version

Evidence:

2024 policy:
Refund period = 30 days.

Current question:

What is the refund period in 2026?

AI:

The current refund period is 30 days.

The Evidence may be genuine but no longer appropriate for the Claim.

Evidence should match:

entity
time
location
product
version
conditions

of the Claim.

44. What Is Citation Hallucination?

A Citation Hallucination occurs when an AI generates a citation or attribution that does not correspond correctly to real supporting information.

One obvious form is:

AI cites a source
that does not exist.

Another form occurs when a real Source is presented as if it supports a Claim that it does not.

For example:

Claim:
Enterprise refunds are available for 30 days.

Citation:
Enterprise Billing Policy

But the cited policy actually says:

14 days

The presence of a citation does not make the Claim supported.

45. Citation Hallucination vs. Unsupported Claim

These failures can occur together.

AI Claim:
Refund period = 30 days

Evidence:
Refund period = 14 days

Citation:
Enterprise Refund Policy

The Claim is:

Unsupported / Contradicted

and the citation is also misleading because it appears to support the Claim when it does not.

But they are conceptually distinct:

Unsupported Claim
→ problem with Claim support

Citation Hallucination
→ problem with citation / attribution

Citation quality will be studied more deeply in Chapter 21.

46. A Citation Does Not Create Grounding

Suppose an answer says:

The refund period is 30 days. [Source 1]

If Source 1 says:

Refund period = 14 days

the answer is not grounded merely because a citation appears next to it.

Citation Presence
≠
Groundedness

Similarly:

Source Mention
≠
Evidence Support

The underlying Claim–Evidence relationship still matters.

47. Hallucination Can Occur Despite Successful Retrieval

Correct Source Retrieved
↓
Correct Passage Selected
↓
Correct Evidence Entered Context
↓
LLM Generates Unsupported Claim

This shows:

Successful Retrieval
≠
Hallucination-Free Generation

Hallucination can occur after retrieval.

This is why RAG does not automatically eliminate hallucinations.

48. RAG Reduces Some Risks but Does Not Eliminate Them

RAG can provide:

fresh information
private information
supporting Evidence
source provenance

But the model can still:

ignore Evidence
misinterpret Evidence
combine Evidence incorrectly
overstate Evidence
attach Evidence to the wrong entity
generate unsupported details
generate incorrect citations

RAG
≠
Guaranteed Grounding

A better model is:

RAG
→ creates conditions for grounded generation

but

Generation still needs
correct Evidence use

49. A Complete Grounding Analysis Example

Suppose the system retrieves:

Evidence A:
Enterprise customers may disable future renewal
from Billing Settings.

Evidence B:
Enterprise annual subscriptions may be refunded
within 14 days of purchase.

Evidence C:
After 14 days, cancellation does not automatically
create refund eligibility.

The AI answers:

Enterprise customers can cancel renewal
through Billing Settings.

They are guaranteed a full refund within 14 days.

After 14 days, cancellation still provides
a partial refund.

Step 1 — Extract the Claims

Claim 1:
Enterprise customers can cancel renewal
through Billing Settings.

Claim 2:
Enterprise customers are guaranteed a full refund
within 14 days.

Claim 3:
After 14 days, cancellation provides
a partial refund.

Step 2 — Claim 1 Groundedness

Evidence A:

Enterprise customers may disable future renewal
from Billing Settings.

Claim 1 closely reflects that Evidence.

Grounded
Faithful

Step 3 — Claim 2 Groundedness

Evidence B:

Enterprise annual subscriptions may be refunded
within 14 days.

Claim 2 says:

guaranteed
full refund

Those details are not established.

Not Fully Grounded

The answer also fails Evidence Faithfulness by overstating the Evidence.

Step 4 — Claim 3 Groundedness

Evidence C:

After 14 days, cancellation does not automatically
create refund eligibility.

Claim 3 says:

cancellation provides a partial refund

This is incompatible with the Evidence.

Contradicted
Ungrounded
Unfaithful

Step 5 — Factuality

To determine Factuality, we need to know the real policy.

The supplied Evidence alone lets us evaluate:

Groundedness
Faithfulness

but external truth verification may be required for:

Factuality

Step 6 — Overall Answer

Claim 1
→ Grounded + Faithful

Claim 2
→ Partially Unsupported + Unfaithful

Claim 3
→ Contradicted + Ungrounded + Unfaithful

The answer as a whole is therefore not fully grounded or faithful.

50. A Complete Three-Dimension Example

Consider the following Evidence:

Outdated Source:
The current subscription price is €100.

Suppose the actual current price is:

€120

The AI answers:

The current subscription price is €100.

Groundedness

Does the available Evidence support the Claim?

Yes.

Faithfulness

Does the answer accurately represent the supplied Evidence?

Yes.

Factuality

Is the statement actually true now?

No.

Therefore:

Grounded
+
Faithful
+
Factually Incorrect

This example shows why the three dimensions must remain separate.

51. Another Three-Dimension Example

Evidence:

The subscription costs €120.

AI:

The subscription costs €120
and includes unlimited storage.

Assume both statements happen to be true in reality.

Claim 1:

€120
→ Grounded
→ Faithful
→ Factually Correct

Claim 2:

unlimited storage
→ Not supported by supplied Evidence
→ Ungrounded relative to this Evidence
→ Potentially Factually Correct

Factually Correct
≠
Automatically Grounded

52. Why This Matters for RAG

RAG systems are sometimes evaluated as if retrieval alone solves factual generation.

But the full path is:

Retrieval
↓
Evidence
↓
Generation
↓
Claims
↓
Groundedness
↓
Faithfulness
↓
Factuality

Each stage introduces different possible failures.

Bad Retrieval
→ wrong Evidence

Good Retrieval + Bad Generation
→ unfaithful Claims

Good Evidence + Outdated Source
→ grounded but factually wrong answer

RAG quality is therefore multi-dimensional.

53. Why This Matters for AI Visibility

AI Visibility is not only about whether a brand or Source is retrieved.

Brand Source
↓
Retrieved
↓
Passage Selected
↓
Evidence
↓
AI Claim About Brand

The final Claim can represent the brand:

accurately
incompletely
incorrectly
too broadly
with unsupported attributes

For example, a Source says:

Company X offers Persian-speaking
accounting services in Amsterdam.

A grounded and faithful Claim might be:

Company X offers Persian-speaking
accounting services in Amsterdam.

An unfaithful Claim might be:

Company X is the largest Persian-speaking
accounting firm in the Netherlands.

The Source may have achieved retrieval visibility.

But the generated representation is no longer faithful to the Evidence.

AI Visibility therefore includes not only:

Was the brand retrieved?

but eventually also:

How was the brand represented
in the generated Claims?

54. Common Misunderstandings

Misunderstanding 1: Grounding means running a fact-check after generation

Incorrect. Grounding means keeping generated Claims tied to Evidence. Post-generation checking is Groundedness Evaluation.

Misunderstanding 2: If Evidence is in the Context, the Answer is Grounded

Incorrect. The model can ignore or contradict Evidence that is present.

Misunderstanding 3: Retrieval guarantees Grounding

Incorrect. Retrieval finds information. Grounding concerns whether generated Claims remain supported by it.

Misunderstanding 4: Groundedness and Factuality are identical

Incorrect. Groundedness is Evidence-relative. Factuality concerns external truth.

Misunderstanding 5: A Grounded Answer must be factually correct

Not necessarily. If the supplied Evidence is wrong or outdated, the answer can be grounded but factually incorrect.

Misunderstanding 6: A Factually Correct Claim must be Grounded

Incorrect. A Claim may be true but unsupported by the Evidence supplied to the model.

Misunderstanding 7: Faithfulness means copying the Source exactly

Incorrect. A paraphrase can be faithful if it preserves the Evidence's meaning.

Misunderstanding 8: Similar wording guarantees Faithfulness

Incorrect. A small change in quantity, certainty, entity, scope, or condition can change the meaning.

Misunderstanding 9: Groundedness and Faithfulness are completely unrelated

Incorrect. They overlap strongly, but emphasize support and accurate representation respectively.

Misunderstanding 10: More Evidence automatically produces a more Grounded Answer

Incorrect. The model can still misuse or misinterpret additional Evidence.

Misunderstanding 11: An Unsupported Claim must be false

Incorrect. It may be true but unsupported by the available Evidence.

Misunderstanding 12: Source Misinterpretation is a Retrieval failure

Not necessarily. The correct Source may have been retrieved and then interpreted incorrectly.

Misunderstanding 13: Evidence Misuse means the Evidence itself is false

Incorrect. The Evidence can be valid but used to support an inappropriate Claim.

Misunderstanding 14: A Citation proves that a Claim is Grounded

Incorrect. The cited Source must actually support the Claim.

Misunderstanding 15: RAG eliminates hallucinations

Incorrect. RAG supplies external Evidence but does not guarantee that the model will use it correctly.

55. How the Concepts Connect

Chapter 18 established:

Passage
↓
Evidence

Chapter 19 established:

Evidence
↓
Claim

and:

Evidence
+
Claim
↓
Entailment
Contradiction
Neutral

Chapter 20 now evaluates the broader relationship:

Evidence
↓
Generated Claims
↓
Groundedness
Faithfulness
Factuality

The larger architecture is:

SOURCE LAYER

Source
↓
Document
↓
Passage


RETRIEVAL LAYER

↓
Retrieved Passage


CONTEXT LAYER

↓
Relevant Context


EVIDENCE LAYER

↓
Evidence


GENERATION LAYER

↓
LLM
↓
Claims


SUPPORT LAYER

Evidence
↕
Claims
↓
Entailment / Contradiction / Neutral


ANSWER QUALITY

↓
Groundedness
Faithfulness
Factuality

The three core questions are now:

Groundedness:
Is the answer supported by Evidence?

Faithfulness:
Does the answer accurately reflect Evidence?

Factuality:
Is the answer actually true?

The next major issue is:

Even if a Claim is supported by Evidence, how should the system communicate where that support came from?

That leads directly into:

Chapter 21 — Citations and Attribution

Mental Model

Imagine a journalist receives a company document saying:

Refund requests must be submitted
within 14 days of purchase.

The journalist writes:

Customers have 14 days
to request a refund.

The statement is supported by the document.

Grounded

The journalist preserved the meaning of the Source.

Faithful

If the actual company policy really is 14 days:

Factually Correct

Now imagine the Source is outdated.

It says:

30 days

but the actual current policy is:

14 days

The journalist accurately reports:

The refund period is 30 days.

Relative to the Source:

Grounded
Faithful

Relative to reality:

Factually Incorrect

Now imagine the Source says:

Some customers may receive refunds.

The journalist writes:

Every customer receives a refund.

The answer is:

Unfaithful
Ungrounded

because the Claim goes beyond the Evidence.

Finally, imagine the journalist attaches a citation to a document that never said what the article claims.

The citation exists.

But the support does not.

That is the beginning of:

Citation Hallucination

The simplest mental model is:

Evidence
↓
AI Claim
↓
Ask Three Questions


1. Groundedness
Does Evidence support the Claim?


2. Faithfulness
Does the Claim preserve the Evidence's meaning?


3. Factuality
Is the Claim actually true?

And the simplest reliable distinction is:

Groundedness asks whether an answer is supported by the available Evidence. Faithfulness asks whether the answer accurately represents that Evidence. Factuality asks whether the answer is actually true in the real world. A high-quality RAG answer should ideally satisfy all three, but one does not automatically guarantee the others.

Chapter 20 Recap

Grounding

Keeping generated Claims tied to available Evidence.

Evidence
↓
Generation
↓
Supported Claims

Grounding vs. Groundedness Evaluation

Grounding
→ generated answer stays supported by Evidence

Groundedness Evaluation
→ check whether that actually happened

Grounding is not inherently a separate post-generation verification stage.

Groundedness

The degree to which Claims in an answer are supported by available Evidence.

Answer
↓
Claims
↓
Evidence Support
↓
Groundedness

Grounded Answer

An answer whose important factual Claims are supported by available Evidence.

Evidence
↓
Supported Claims
↓
Grounded Answer

Evidence in Context vs. Grounding

Evidence in Context
≠
Automatically Grounded Answer

The model must use the Evidence correctly.

Faithfulness

Whether the generated answer accurately represents the supplied Evidence.

Evidence
↓
Generated Representation
↓
Faithfulness

Evidence Faithfulness

Whether Claims preserve the meaning, conditions, scope, and certainty of the Evidence.

Evidence Meaning
≈
Claim Meaning

Groundedness vs. Faithfulness

Groundedness
→ Is the Claim supported?

Faithfulness
→ Does the Claim accurately reflect the Evidence?

They overlap but emphasize different aspects of answer quality.

Factuality

Whether the Claim is actually true.

Claim
↓
Compare With Reality
↓
True / False

Factual Accuracy

The accuracy of factual statements relative to real-world truth or reliable ground truth.

Groundedness vs. Factuality

Groundedness
→ supported by available Evidence

Factuality
→ actually true

These are different dimensions.

Grounded but Factually Incorrect

Evidence is wrong / outdated
+
Answer accurately follows Evidence

Result:

Grounded
Faithful
Factually Incorrect

Factually Correct but Ungrounded

Claim happens to be true
but
supplied Evidence does not support it

Result:

Factually Correct
Ungrounded

Unsupported Claim

A Claim that lacks adequate Evidence support.

Claim
↓
No Adequate Evidence
↓
Unsupported

Unsupported does not necessarily mean false.

Source Misinterpretation

The model misunderstands what the Evidence says.

Correct Evidence
↓
Incorrect Interpretation
↓
Incorrect Claim

Evidence Misuse

Real Evidence is used to justify a Claim it does not support.

Valid Evidence
+
Invalid Inference
↓
Evidence Misuse

Citation Hallucination

A citation or attribution is generated that does not correctly correspond to real supporting information.

Claim
+
Citation
↓
Citation does not actually support Claim

Citation Presence vs. Grounding

Citation Present
≠
Claim Grounded

The cited Source must actually support the Claim.

RAG vs. Hallucination

RAG
→ provides external Evidence

but

RAG
≠
guaranteed correct Evidence use

Hallucination can still occur after successful Retrieval.

The Three Core Evaluation Questions

1. Groundedness
Is the Claim supported by Evidence?

2. Faithfulness
Does the Claim accurately reflect Evidence?

3. Factuality
Is the Claim actually true?

The complete Chapter 20 pipeline is:

Corpus
↓
Retrieval
↓
Passage
↓
Evidence
↓
LLM Generation
↓
Claims
↓
Claim–Evidence Analysis
↓
Groundedness
↓
Faithfulness
↓
Factuality

Possible generation failures include:

Unsupported Claim
Source Misinterpretation
Evidence Misuse
Citation Hallucination

The critical course pipeline has now become:

Source
↓
Passage
↓
Evidence
↓
Claim
↓
Groundedness / Faithfulness / Factuality

The next chapter is:

Chapter 21 — Citations and Attribution

If Evidence supports a Claim, how should the system connect that Claim back to its Source so the user can see where the information came from?

The Single Most Important Idea from Chapter 20

Groundedness, Faithfulness, and Factuality evaluate different relationships. Groundedness asks whether generated Claims are supported by the available Evidence. Faithfulness asks whether the answer accurately reflects that Evidence without distortion or unsupported additions. Factuality asks whether the Claims are actually true in the real world. A RAG answer can be grounded and faithful yet factually wrong if its Evidence is wrong or outdated, and it can be factually correct yet ungrounded if it states true information that is not supported by the supplied Evidence. High-quality AI generation ideally requires all three.