Chapter 18- From Source to Evidence

Chapter Goal

In the previous chapters, we built the retrieval pipeline from the searchable collection to the final LLM context.

At a high level:

Sources
↓
Documents
↓
Parsing / Chunking / Indexing
↓
Corpus / Search Index

        QUERY TIME

User Query
↓
Retrieval
↓
Retrieved Documents / Passages
↓
Context Selection
↓
Final LLM Context

Chapter 18 focuses on a new question:

When does information from a retrieved passage become evidence for an answer?

Before answering that question, one distinction must remain clear.

Corpus
≠
Source

The Corpus is the searchable collection.

A Source is the origin of a document inside that collection.

For example:

Corpus / Search Index

├── Document A
│   Source: company.com
│
├── Document B
│   Source: company.com
│
├── Document C
│   Source: government.nl
│
└── Document D
    Source: research-journal.org

Retrieval searches the Corpus and returns Documents or Passages.

User Query
↓
Retriever
↓
Corpus
↓
Retrieved Document / Passage

Each retrieved item still has a Source.

Retrieval does not transform the Corpus into a Source.

The Source was already the origin of the retrieved information.

Chapter 17 also introduced this distinction:

Relevant Context
≠
Automatically Evidence

A passage may be useful background information without directly supporting what the answer says.

This chapter focuses on the transition:

Source
↓
Source Document
↓
Passage
↓
Relevant Information
↓
Evidence

To understand Evidence, we also need one concept that will be studied formally in Chapter 19:

A Claim is a factual statement made in the AI's answer that can potentially be supported or contradicted by evidence.

For example:

AI Answer:
The Enterprise refund period is 14 days.

Claim:
The Enterprise refund period is 14 days.

Evidence might be:

Enterprise customers may request a refund
within 14 days of the original purchase.

The relationship is:

Evidence
↓
supports
↓
Claim

This chapter does not yet analyze Claims in depth. That is the purpose of Chapter 19.

Here, we only need to understand what information can function as evidence.

By the end of this chapter, you should understand:

how Source relates to Corpus;
what a Source is;
what a Source Document is;
what a Passage is in the source-to-evidence pipeline;
when a passage becomes Evidence;
what Supporting Evidence means;
what Evidence Retrieval means;
what Evidence Coverage is;
what Relevant Evidence means;
what Sufficient Evidence means;
what Conflicting Evidence is;
and why evidence quality matters.

18.1 Source

1. Corpus and Source

Before defining Source, it is important to connect this chapter to the retrieval concepts from earlier chapters.

A Corpus is the searchable collection available to the retrieval system.

It may contain documents originating from many different sources.

Corpus

├── Document 1 ← Source A
├── Document 2 ← Source A
├── Document 3 ← Source B
├── Document 4 ← Source C
└── Document 5 ← Source C

The relationship is therefore not:

Corpus
↓
Retrieval
↓
Source

Instead:

Sources
↓
Documents
↓
Processing / Indexing
↓
Corpus / Search Index

Then, at query time:

User Query
↓
Retriever
↓
Corpus / Search Index
↓
Retrieved Document / Passage

Each retrieved Document or Passage still belongs to, or originates from, a Source.

Retrieved Passage
↓
came from
↓
Source Document
↓
came from
↓
Source

So:

Corpus
→ searchable collection

Source
→ origin of information

Retrieval
→ selects Documents / Passages from the Corpus

2. What Is a Source?

A Source is the origin or provider from which information comes.

Examples can include:

website
government database
company knowledge base
news publisher
research repository
internal documentation system

For example:

Source:
Company Help Center

may contain hundreds of separate documents.

Source
→ information origin or provider

3. Source vs. Information

A Source is not the same thing as the individual information extracted from it.

For example:

Source:
Government Website

may contain information about:

citizenship
residence permits
taxes
business registration
health insurance

The Source identifies where the information comes from.

The individual facts exist inside documents and passages from that Source.

Source
↓
Documents
↓
Passages
↓
Information

4. What Is a Source Document?

A Source Document is a specific document obtained from a Source.

For example:

Source:
Company Help Center

Source Document:
Annual Subscription Refund Policy

Or:

Source:
Government Website

Source Document:
Naturalisation Requirements

Source
→ where the information comes from

Source Document
→ specific document containing the information

5. One Source Can Contain Many Source Documents

A single Source may contain many individual documents.

Source:
Company Knowledge Base

├── Billing Policy
├── Refund Policy
├── Cancellation Policy
├── Account Security Guide
└── Privacy Policy

These documents may all become part of the searchable Corpus.

One Source
↓
Many Documents
↓
Corpus / Search Index

6. Source Document vs. Retrieved Document

A document can exist in the Corpus without being retrieved for a particular query.

Corpus

├── Document A
├── Document B
├── Document C
└── Document D

A query might retrieve:

Document B
Document D

So:

Source Document
→ document available from a Source

Retrieved Document
→ document selected during retrieval
   for the current query

Retrieval changes the document's role in the current interaction, not its original Source.

7. What Is a Passage?

A Passage is a smaller coherent portion of a Source Document.

For example:

Source:
Company Help Center

Source Document:
Annual Subscription Policy

Passage:
Annual subscriptions may be refunded
within 14 days of the original purchase.

The hierarchy is:

Source
↓
Source Document
↓
Passage

The Passage is usually closer to the specific information needed for answering a query than the entire document.

8. One Document Can Produce Many Passages

Suppose a document contains sections about:

billing
renewal
cancellation
refunds
exceptions

These may correspond to:

Passage A → billing
Passage B → renewal
Passage C → cancellation
Passage D → refunds
Passage E → exceptions

A refund query may only need:

Passage D
Passage E

This is why evidence work often happens at the passage level rather than at the level of the entire document.

9. Source Is Not Yet Evidence

Suppose the user asks:

What is the Enterprise refund period?

We may know the Source:

Company Help Center

But that does not yet answer the question.

We may know the Source Document:

Enterprise Billing Policy

But the document may contain many unrelated sections.

The useful Passage is:

Enterprise customers may request a refund
within 14 days of purchase.

That Passage can support:

The Enterprise refund period is 14 days.

Source
↓
Source Document
↓
Passage
↓
Potential Evidence

The Source provides origin and provenance.

The Passage provides specific supportable information.

18.2 Evidence

10. What Is Evidence?

Evidence is information that can support a statement, answer, or Claim.

Suppose the question is:

How long is the Enterprise refund period?

Passage:

Enterprise customers may request a refund
within 14 days of the original purchase.

This Passage can function as Evidence for:

The Enterprise refund period is 14 days.

Evidence
↓
supports
↓
Answer / Claim

Evidence is therefore defined partly by a relationship.

Information becomes evidence because it supports something the answer needs to state.

11. What Is a Claim?

Claim will be studied formally in Chapter 19.

For this chapter, we only need a simple definition:

A Claim is a factual statement made in the AI's answer that can potentially be supported or contradicted by evidence.

For example:

AI Answer:
The refund period is 14 days.

The statement:

The refund period is 14 days.

is a Claim.

If the retrieved Passage says:

Enterprise customers may request a refund
within 14 days of purchase.

then:

Evidence
↓
supports
↓
Claim

This relationship will become the central topic of Chapter 19.

12. Passage vs. Evidence

A Passage is a unit of source content.

Evidence is the role that information plays when it supports an answer or Claim.

Consider:

Passage A:
Enterprise plans include priority support.

Passage B:
Enterprise plans are refundable within 14 days.

Question:

What is the refund period?

Both A and B are Passages.

But:

Passage A
→ not Evidence for the refund period

Passage B
→ Evidence for the refund period

Passage
→ information unit

Evidence
→ information used as support

13. Evidence Is Query- and Claim-Dependent

Consider this Passage:

Enterprise annual plans cost €500
and may be refunded within 14 days.

For:

How much does Enterprise cost?

the relevant Evidence is:

Enterprise annual plans cost €500.

For:

What is the refund period?

the relevant Evidence is:

may be refunded within 14 days.

Evidence
depends on
what needs to be supported

The same Passage can provide different evidence for different questions or Claims.

14. What Is Supporting Evidence?

Supporting Evidence is evidence that directly supports a specific answer or Claim.

Claim:

Enterprise refunds are available within 14 days.

Supporting Evidence:

Enterprise customers may request a refund
within 14 days of the original purchase.

Supporting Evidence
↓
supports
↓
Claim

15. Evidence Can Be Smaller Than the Retrieved Passage

Suppose the retrieved Passage says:

Enterprise plans renew annually.
Customers may cancel renewal from Billing Settings.
Refund requests are accepted within 14 days
of the original annual purchase.
Priority support is available seven days a week.

The question is:

What is the refund period?

The entire Passage may be available in context.

But the specific Evidence is:

Refund requests are accepted within 14 days
of the original annual purchase.

Retrieved Passage
↓
Relevant Context
↓
Specific Supporting Information
↓
Evidence

16. Context vs. Evidence Revisited

From Chapter 17:

Context
→ information available to the LLM

Evidence
→ information that supports what the answer says

Consider:

Context:

A:
Enterprise is the premium plan.

B:
Enterprise subscriptions renew annually.

C:
Refunds are available within 14 days.

Question:

How long is the refund period?

All three may be available to the model.

But direct Evidence is:

So a useful conceptual relationship is:

Relevant Context
↓
Supporting Information
↓
Evidence

Evidence is often a more specific subset of the context.

17. What Is Evidence Retrieval?

Evidence Retrieval means retrieving information with the goal of finding material capable of supporting the answer or its Claims.

Instead of asking only:

Which documents are related to this topic?

the system is interested in:

Which passages contain information
that can support the answer?

Question
↓
Retrieval
↓
Candidate Passages
↓
Potential Supporting Evidence

18. Evidence Retrieval vs. General Retrieval

General Retrieval may find:

topically related information

Evidence Retrieval emphasizes:

supporting information

Example:

Question:
Does the Enterprise plan have a 14-day refund period?

General relevant Passage:

Enterprise is the company's premium subscription tier.

Evidence Passage:

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

Both are related to Enterprise.

Only the second directly supports the answer.

Relevance
is necessary

but

Relevance alone
may not be sufficient
for Evidence

19. Evidence Retrieval May Reuse Earlier Retrieval Results

Evidence Retrieval does not necessarily require another completely separate search.

A simple architecture may look like:

Initial Retrieval
↓
Retrieved Passages
↓
Context Selection
↓
Identify Supporting Evidence

In another system:

Initial Retrieval
↓
Insufficient Support
↓
Additional Retrieval
↓
More Evidence Candidates

So Evidence Retrieval describes the goal:

find supporting information

not one mandatory implementation pattern.

20. What Is Evidence Coverage?

Evidence Coverage describes how much of the information that needs support is actually covered by available evidence.

Suppose the user asks:

Can I cancel my plan,
when does cancellation take effect,
and can I receive a refund?

The answer requires three information components:

1. cancellation process
2. effective date
3. refund eligibility

Suppose available evidence supports:

1. cancellation process
2. effective date

but nothing supports:

3. refund eligibility

Evidence Coverage is therefore incomplete.

21. Complete vs. Incomplete Evidence Coverage

Suppose the required answer contains:

A
B
C

Available evidence provides:

Evidence 1 → supports A
Evidence 2 → supports B

Nothing supports C.

Evidence Coverage

A → covered
B → covered
C → not covered

A system may therefore have strong evidence for some parts of an answer but not all of them.

22. Evidence Coverage vs. Retrieval Recall

These concepts are related but different.

Retrieval Recall

Did retrieval find the relevant items?

Evidence Coverage

Do we have supporting information for everything the answer needs to state?

A system can retrieve many relevant documents and still lack evidence for an important part of the answer.

High Retrieval Recall
≠
Complete Evidence Coverage

23. Multiple Pieces of Evidence May Be Needed

Some questions require several independent pieces of information.

Question:

Which plan is cheaper,
and which plan has the longer refund period?

The answer may require:

Passage A
→ Standard price

Passage B
→ Enterprise price

Passage C
→ Standard refund period

Passage D
→ Enterprise refund period

Complex Answer
↓
Multiple Information Needs
↓
Multiple Supporting Evidence Pieces

No single Passage necessarily covers the entire answer.

18.3 Evidence Quality

24. What Is Evidence Quality?

Not all evidence is equally useful.

A Passage may appear relevant but still be:

weak
incomplete
outdated
ambiguous
conflicting

Evidence Quality describes how suitable the available evidence is for supporting the required answer.

In this chapter, the core concepts are:

Relevant Evidence
Sufficient Evidence
Conflicting Evidence

25. What Is Relevant Evidence?

Relevant Evidence directly relates to the specific information that needs support.

Question:

What is the Enterprise refund period?

Passage A:

Enterprise plans include priority support.

Passage B:

Enterprise annual subscriptions may be refunded
within 14 days.

Passage B is Relevant Evidence for the refund-period answer.

Relevant Evidence
→ directly connected to what needs support

26. Topical Relevance Is Not Enough

Suppose the question is:

Does cancellation automatically create refund eligibility?

Passage:

Enterprise customers can cancel from Billing Settings.

This Passage is topically related to cancellation.

But it does not answer:

refund eligibility

A stronger Passage is:

Cancellation after the refund period does not
automatically create refund eligibility.

Topically Related
≠
Directly Supporting

27. What Is Sufficient Evidence?

Sufficient Evidence contains enough information to support the required answer or Claim.

Question:

Can I receive a refund after cancelling?

Evidence A:

Customers can cancel from Billing Settings.

This is not sufficient.

It tells us about cancellation but not refund eligibility.

Evidence B:

Refunds are available within 14 days.
Cancellation after that period does not
automatically create refund eligibility.

This provides much stronger support for the requested answer.

Relevant
≠
Always Sufficient

28. Relevant Evidence vs. Sufficient Evidence

Relevant Evidence

related to the information
that needs support

Sufficient Evidence

contains enough information
to support what needs to be stated

Example:

Question:
Are refunds available after cancellation,
and what is the deadline?

Evidence:

Refunds are available.

This is relevant.

But it is not sufficient because it does not provide the deadline.

Relevant
→ points in the right direction

Sufficient
→ provides enough support

29. Sufficiency Depends on the Claim

The amount and precision of evidence required depends on what the answer claims.

Claim:

Refunds are available.

Evidence:

Customers may request refunds.

may be sufficient.

But consider a more specific Claim:

Enterprise customers can receive a full refund
within 14 days regardless of usage.

Now the Evidence needs to support:

Enterprise customers
full refund
14 days
regardless of usage

More Specific Claim
↓
More Specific Evidence Requirements

The detailed relationship between Claims and Evidence will be studied in Chapter 19.

30. Evidence Can Support Only Part of a Statement

Suppose the answer says:

Enterprise customers receive a full refund
within 14 days.

Available Evidence says:

Enterprise customers may request a refund
within 14 days.

The Evidence supports:

Enterprise customers
refund
14 days

But it does not necessarily support:

full refund

Conceptually:

Claim

├── Enterprise customers → supported
├── refund               → supported
├── 14 days              → supported
└── full refund          → not established

This distinction becomes central in the next chapter.

31. What Is Conflicting Evidence?

Conflicting Evidence occurs when different evidence sources or passages support incompatible conclusions.

For example:

Source A:
Refund period = 14 days

Source B:
Refund period = 30 days

Evidence A
↔
Evidence B

If both refer to the same policy under the same conditions, they conflict.

32. Why Conflicting Evidence Matters

Suppose the LLM receives:

Passage A:
Refunds are available within 14 days.

Passage B:
Refunds are available within 30 days.

The model cannot safely assume that both statements describe the same current policy.

The difference might exist because:

one source is outdated
different products are involved
different countries are involved
the policy changed over time
one source is incorrect

Additional context may be required before the conflict can be resolved.

33. Conflict Can Be Real or Apparent

Sometimes two Passages appear contradictory but describe different conditions.

Example:

Passage A:
Standard plans have a 30-day refund period.

Passage B:
Enterprise plans have a 14-day refund period.

These do not conflict.

The difference is explained by:

plan type

But if both say:

Enterprise refund period

with incompatible values under the same conditions, there may be a genuine conflict.

Different Information
≠
Automatically Conflicting Evidence

34. Metadata Can Help Explain Conflicts

Suppose:

Source A
updated = 2025
refund period = 30 days

Source B
updated = 2026
refund period = 14 days

The metadata suggests that the policy may have changed.

Or:

Source A
country = Canada

Source B
country = Netherlands

The difference may be geographic.

Useful metadata may include:

date
location
product
version
source identity

These attributes can help determine whether evidence truly conflicts.

35. More Evidence Does Not Automatically Mean Better Evidence

Imagine one retrieval result contains:

10 weak passages

that vaguely relate to the question.

Another contains:

2 highly relevant and sufficient passages

The second evidence set may be much more useful.

Evidence Quantity
≠
Evidence Quality

The objective is not merely to collect more Passages.

Useful evidence should be:

relevant
sufficient
usable
consistent enough for the task

36. A Complete Source-to-Evidence Example

Suppose the question is:

Can an Enterprise customer receive a refund
after cancelling an annual subscription?

Step 1 — Corpus

The retrieval system searches a Corpus containing many indexed documents from multiple Sources.

Corpus

├── Company Help Center documents
├── Product documentation
├── Billing documentation
└── Other indexed content

Step 2 — Retrieval

The Retriever searches the Corpus and returns a relevant document and its passages.

User Query
↓
Retriever
↓
Corpus
↓
Retrieved Document / Passages

Step 3 — Source

The retrieved information originates from:

Source:
Company Help Center

Step 4 — Source Document

The specific retrieved document is:

Enterprise Billing and Refund Policy

Step 5 — Passages

The document contains:

Passage A:
Enterprise subscriptions renew annually.

Passage B:
Customers can cancel future renewal
from Billing Settings.

Passage C:
Enterprise annual subscriptions may be refunded
within 14 days of the original purchase.

Passage D:
After the 14-day refund period,
cancellation prevents future renewal but does not
automatically create refund eligibility.

Passage E:
Enterprise customers receive priority support.

Step 6 — Relevant Context

For the question, useful context is:

B
C
D

Step 7 — Evidence

Passage B supports:

Cancellation is possible.

Passage C supports:

Refunds are available within 14 days.

Passage D supports:

Cancellation after 14 days does not
automatically create refund eligibility.

These Passages now function as Evidence because they support the information required in the answer.

Step 8 — Evidence Coverage

The question contains two main information needs:

Can I cancel?

Can I receive a refund after cancellation?

The evidence covers both:

Cancellation
→ Passage B

Refund eligibility
→ Passages C + D

Step 9 — Evidence Quality

Relevant?

Yes.

Sufficient for the basic answer?

Yes.

Visible conflict?

No.

The system therefore has a useful evidence set for generation.

37. The Full Source-to-Evidence Pipeline

The complete architecture should preserve the Corpus and Retrieval stages from earlier chapters.

SOURCES

Source A
Source B
Source C
↓
Documents
↓
Parsing / Chunking / Indexing
↓
Corpus / Search Index


QUERY TIME

User Query
↓
Retrieval
↓
Retrieved Document / Passage
↓
Context Selection
↓
Relevant Passage
↓
Evidence
↓
Supporting Evidence
↓
Evidence Coverage
↓
Evidence Quality

A more compact view is:

Corpus
↓
Retrieve
↓
Document / Passage
↓
Identify Source
↓
Select Relevant Information
↓
Evidence

And the support relationship is:

Evidence
↓
supports
↓
Claim in the AI Answer

38. Evidence Is a Role, Not a Special File Type

There is no special content format called:

Evidence

A sentence or Passage becomes Evidence because of its relationship to what needs to be supported.

For example:

Refunds are available within 14 days.

For:

What is the refund period?

it is Evidence.

For:

What is the subscription price?

it is not Evidence.

Evidence
=
Information
+
Support Relationship

39. Why This Matters for RAG

A basic RAG pipeline is often simplified as:

Retrieve
↓
Context
↓
Generate

But high-quality grounded generation requires more than topical context.

A more useful conceptual model is:

Retrieve
↓
Relevant Context
↓
Supporting Evidence
↓
Generate Claims

The quality of the final answer therefore depends partly on whether the information shown to the model actually supports what it says.

40. Why This Matters for AI Visibility

From an AI Visibility perspective, being present in the searchable Corpus is only the beginning.

A source may move through several stages:

Source Exists
↓
Document Is Indexed
↓
Document / Passage Is Retrieved
↓
Passage Is Selected Into Context
↓
Passage Functions as Evidence
↓
Evidence Supports the Answer

So:

Indexed
≠
Retrieved

Retrieved
≠
Selected Into Context

Context
≠
Automatically Evidence

A webpage may be retrieved but contain no passage that sufficiently supports the required answer.

Another webpage may contain a clear, precise supporting passage.

The second is more useful as evidence.

41. Common Misunderstandings

Misunderstanding 1: Source replaces Corpus in the pipeline

Incorrect. The Corpus is the searchable collection. Sources are the origins of documents inside that collection.

Misunderstanding 2: Retrieval happens between Corpus and Source

Incorrect. Retrieval searches the Corpus and returns Documents or Passages. Those items already have Sources.

Misunderstanding 3: Retrieval returns a Source as a new information object

More precisely, retrieval usually returns a Document or Passage from a Source.

Misunderstanding 4: Source and Source Document are the same thing

Not necessarily. One Source can contain many Source Documents.

Misunderstanding 5: Every Passage from a trustworthy Source is Evidence

Incorrect. A Passage becomes Evidence only when it supports what the answer needs to state.

Misunderstanding 6: Relevant information and Evidence are always identical

Incorrect. Information may be useful context without directly supporting a Claim.

Misunderstanding 7: Evidence must come from a separate retrieval stage

Incorrect. Evidence may already exist inside Passages retrieved during the original retrieval process.

Misunderstanding 8: Evidence Retrieval always means searching the Corpus again

Incorrect. It describes the goal of finding supporting information and may reuse existing retrieval results.

Misunderstanding 9: A Claim is something the Source claims

In the pipeline studied here, Claim usually refers to a factual statement made in the AI's answer. Source statements may provide Evidence for that Claim.

Misunderstanding 10: One relevant Passage is always sufficient Evidence

Incorrect. It may support only one part of the required answer.

Misunderstanding 11: Relevant Evidence is automatically Sufficient Evidence

Incorrect. Evidence can be relevant but incomplete.

Misunderstanding 12: Evidence Coverage is the same as Retrieval Recall

Incorrect. Retrieval Recall concerns finding relevant items. Evidence Coverage concerns whether the required answer has enough supporting information.

Misunderstanding 13: More Evidence automatically means stronger Evidence

Incorrect. Quality and sufficiency matter more than raw quantity.

Misunderstanding 14: Different values always indicate conflicting Evidence

Not necessarily. The Passages may refer to different dates, products, locations, or conditions.

Misunderstanding 15: Evidence is automatically true

Incorrect. A Source can itself contain outdated, incomplete, or incorrect information.

Misunderstanding 16: A Passage is permanently Evidence

Incorrect. Its Evidence role depends on what answer or Claim it is being used to support.

42. How the Concepts Connect

Earlier chapters established:

Sources
↓
Documents
↓
Parsing
↓
Chunking
↓
Indexing
↓
Corpus / Search Index

Chapter 16 prepared the query:

User Query
↓
Query Understanding
↓
Query Variants

Retrieval then searched the Corpus:

Query
+
Corpus
↓
Retrieval
↓
Candidate Documents / Passages

Chapter 17 transformed those candidates into context:

Candidate Passages
↓
Context Selection
↓
Filtering
↓
Compression
↓
Final Context

Chapter 18 asks which parts of that information can actually support the answer:

Source
↓
Source Document
↓
Passage
↓
Relevant Context
↓
Evidence
↓
Supporting Evidence

Then we evaluate whether the Evidence is:

Relevant
Sufficient
Conflicting

The larger architecture is now:

SOURCE LAYER

Sources
↓
Documents


INDEX LAYER

Parsing / Chunking / Indexing
↓
Corpus / Search Index


QUERY LAYER

User Query
↓
Query Understanding


RETRIEVAL LAYER

↓
Retrieved Documents / Passages


CONTEXT LAYER

↓
Relevant Context


EVIDENCE LAYER

↓
Supporting Evidence
↓
Evidence Coverage
↓
Evidence Quality


GENERATION LAYER

↓
AI Answer
↓
Claims

We now understand:

Chapter 17
→ What information should the LLM see?

Chapter 18
→ Which information can support what the LLM says?

The next question is:

What exactly is the statement that the Evidence is supposed to support?

That leads directly into:

Chapter 19 — Claims

Mental Model

Imagine a journalist researching a company.

The journalist has access to a large research archive.

That archive is similar to:

Corpus

The archive contains documents from:

company websites
government databases
newspapers
research reports

Those are:

Sources

The journalist searches the archive for:

Company X subscription price increase

That is:

Retrieval

The search returns:

Company X — 2026 Pricing Announcement

That is a:

Source Document

Inside the document, one paragraph says:

The Professional plan will increase
from €100 to €120 per month.

That is a:

Passage

The journalist writes:

The Professional plan increased by 20%.

That sentence is a:

Claim

The paragraph from the pricing announcement can function as:

Evidence

because it supports that Claim.

But another Passage saying:

Company X launched the Professional plan in 2023.

may still be useful background context.

It is not Evidence for the price increase.

So:

Corpus
→ where the search happens

Source
→ where a document came from

Document
→ specific information object

Passage
→ specific section of the document

Evidence
→ information that supports what is being said

Claim
→ the statement being supported

The simplest reliable mental model is:

Corpus
↓
Retrieve
↓
Document / Passage
↓
Relevant Context
↓
Evidence
↓
supports
↓
Claim

And the simplest reliable definition is:

Evidence is not a separate type of content. It is information from a retrieved Source Document or Passage that plays a specific role: supporting something the answer says. The Corpus defines where retrieval searches, the Source identifies where the retrieved information originated, and Evidence describes how that information supports the answer.

Chapter 18 Recap

Corpus

The searchable collection used by the retrieval system.

Corpus
├── Documents from Source A
├── Documents from Source B
└── Documents from Source C

Source

The origin or provider of information.

Source
→ website
→ publisher
→ database
→ knowledge base

Corpus vs. Source

Corpus
→ searchable collection

Source
→ origin of a Document inside that collection

Retrieval

Selects Documents or Passages from the Corpus.

User Query
↓
Retriever
↓
Corpus
↓
Retrieved Document / Passage

Source Document

A specific Document originating from a Source.

Source
↓
Source Document

Passage

A smaller coherent section of a Source Document.

Source
↓
Document
↓
Passage

Claim

A factual statement made in the AI's answer that can potentially be supported or contradicted by Evidence.

AI Answer
↓
Claim

Claims are studied formally in Chapter 19.

Evidence

Information that can support a statement, answer, or Claim.

Evidence
↓
supports
↓
Claim

Passage vs. Evidence

Passage
→ information unit

Evidence
→ information used as support

A Passage is not automatically Evidence.

Supporting Evidence

Evidence directly supporting a specific answer or Claim.

Supporting Evidence
↓
Specific Claim

Context vs. Evidence

Context
→ information available to the model

Evidence
→ information supporting what the model says

Evidence Retrieval

Retrieval focused on finding information capable of supporting the required answer.

Question
↓
Retrieve Supporting Information
↓
Evidence Candidates

General Retrieval vs. Evidence Retrieval

General Retrieval
→ find related information

Evidence Retrieval
→ find supporting information

Evidence Coverage

How much of the information that needs support is covered by available Evidence.

Answer Need A → covered
Answer Need B → covered
Answer Need C → not covered

Relevant Evidence

Evidence directly related to the information that needs support.

Sufficient Evidence

Evidence containing enough information to support what needs to be stated.

Relevant
≠
Always Sufficient

Relevant vs. Sufficient Evidence

Relevant
→ connected to what needs support

Sufficient
→ enough to support it

Conflicting Evidence

Evidence that supports incompatible conclusions under the same relevant conditions.

Evidence A
→ 14 days

Evidence B
→ 30 days

Evidence Quality

In this chapter, Evidence Quality focuses on:

Relevant Evidence
+
Sufficient Evidence
+
Conflicting Evidence

Evidence Is Claim-Dependent

The same Passage can function as Evidence for one Claim and not for another.

Passage
+
Claim A
→ Evidence

Same Passage
+
Claim B
→ Not Evidence

The complete Chapter 18 pipeline is:

Sources
↓
Documents
↓
Parsing / Chunking / Indexing
↓
Corpus / Search Index

        QUERY TIME

User Query
↓
Retrieval
↓
Retrieved Document / Passage
↓
Relevant Context
↓
Evidence
↓
Supporting Evidence
↓
Evidence Coverage
↓
Evidence Quality
├── Relevant Evidence
├── Sufficient Evidence
└── Conflicting Evidence
↓
Generation
↓
Claims

The course pipeline has now reached:

Corpus
↓
Retrieval
↓
Context
↓
Evidence
↓
Claims

The next chapter is:

Chapter 19 — Claims

What exactly is a Claim, and how do we determine whether Evidence really supports it?

The Single Most Important Idea from Chapter 18

The Corpus is the collection the retrieval system searches; a Source is the origin of a Document inside that collection. Retrieval returns Documents or Passages from those Sources. A Passage becomes Evidence only when its information is capable of supporting something the AI answer says. Evidence is therefore not another type of document—it is a support role played by information in relation to an answer or Claim.