Skip to content

Mo Farahani

AI Visibility Guide

02 Part II — Information Retrieval Chapter 02.05

Chapter 11- Ranking and Reranking

By Mo Farahani 22 min read

Chapter Goal

In Chapters 7–10, we focused mainly on finding candidates.

We learned several ways to do that:

Lexical Retrieval
→ exact words and term statistics

Dense Retrieval
→ semantic similarity

Hybrid Retrieval
→ combines both

A simplified pipeline now looks like:

User Query
↓
Candidate Generation
↓
Candidate Set
↓
???
↓
Final Results

The missing question is:

Once we have candidates, how do we decide which ones are actually the best?

Suppose hybrid retrieval produces:

Document A
Document B
Document C
Document D
Document E

Finding these candidates is not enough.

The system must determine:

Which document is most relevant?

Which should appear second?

Which candidate only looks relevant superficially?

Which candidate best satisfies the user's actual intent?

This is the job of ranking.

Sometimes an initial retrieval system already assigns scores and produces an order.

But that first ordering can be improved by a more expensive model that examines a much smaller candidate set.

That second stage is called reranking.

The central pipeline of this chapter is:

Large Corpus
↓
Fast Retrieval
↓
Candidate Set
↓
Initial Ranking
↓
Reranking
↓
Final Ranked Results

By the end of this chapter, you should understand:

  • what ranking is;
  • what a ranker is;
  • what a ranking score is;
  • how ranking differs from retrieval;
  • why retrieval systems use multiple stages;
  • what first-stage ranking means;
  • what reranking is;
  • why reranking can be more computationally expensive;
  • what a cross-encoder is;
  • how cross-encoders differ from bi-encoders;
  • why joint query–document processing can improve relevance estimation;
  • what ranking features are;
  • what Learning to Rank means;
  • what pointwise, pairwise, and listwise ranking mean conceptually;
  • how lexical, semantic, freshness, authority, and other signals can influence ranking;
  • how ranking differs from factual verification;
  • how reranking fits into hybrid retrieval and RAG;
  • and why ranking quality matters for AI Visibility.

1. Candidate Generation Is Not the Final Decision

Recall the hybrid retrieval pipeline:

User Query
├──→ Lexical Retrieval
└──→ Dense Retrieval
↓
Fusion
↓
Candidate Set

Suppose the result is:

A
B
C
D
E
F

These documents are:

potentially relevant

not necessarily:

correctly ordered

Candidate generation answers:

Which documents are worth considering?

Ranking answers:

Which of those documents should come first?

This distinction is fundamental.

2. What Is Ranking?

Ranking is the process of ordering candidate items according to an estimated measure of relevance, usefulness, or another objective.

Candidate Set
↓
Evaluate Candidates
↓
Assign Scores
↓
Sort
↓
Ranked Results

For example:

Candidate Set:
A
B
C
D

After ranking:

1. C
2. A
3. D
4. B

The system has transformed:

unordered or weakly ordered candidates

into:

an ordered result list

3. What Is a Ranker?

A ranker is the component or model responsible for assigning ranking scores or producing an ordering.

Query
+
Candidate Document
↓
Ranker
↓
Relevance Score

For multiple candidates:

Query
+
A
B
C
D
↓
Ranker
↓
A → 0.76
B → 0.31
C → 0.91
D → 0.62

Then:

1. C
2. A
3. D
4. B

The exact meaning of the scores depends on the ranking system.

4. Ranking Score

A ranking score is a numerical value used to help determine result order.

For example:

Document A → 8.7
Document B → 4.2
Document C → 9.5

Therefore:

C > A > B

But remember:

Ranking Score
≠
Probability of Truth

and usually:

Ranking Score
≠
Universal Score

A score of:

9.5

in one ranker cannot automatically be compared with:

9.5

from another ranking system.

5. Retrieval vs. Ranking

Retrieval

Huge Corpus
↓
Find Potential Candidates

Question:

What should we consider?

Ranking

Candidates
↓
Evaluate and Order

Question:

Which candidates are best?

So:

Retrieval
→ reduces the search space

Ranking
→ orders what survived

6. Retrieval Often Already Produces an Initial Ranking

The distinction is conceptually useful, but real systems can overlap.

For example, BM25 retrieval does not merely find documents.

It also assigns scores:

Doc A → BM25 14.2
Doc B → BM25 10.3
Doc C → BM25 6.9

Dense retrieval similarly produces similarity scores:

Doc A → 0.91
Doc B → 0.84
Doc C → 0.66

So retrieval systems often produce an initial ranking.

Retrieval
↓
Candidates + Initial Scores
↓
Initial Ranking

The next question is whether that initial ranking is good enough.

Often, it is not.

7. First-Stage Ranking

The ranking produced during initial retrieval is often called first-stage ranking.

The first stage needs to process a very large search space.

Therefore it usually prioritizes:

speed
scalability
reasonable relevance

rather than maximum precision.

For example:

10,000,000 Documents
↓
BM25 / Dense / Hybrid Retrieval
↓
Top 500 Candidates

That first stage must be fast.

It cannot usually perform an extremely expensive analysis of all ten million documents.

8. Why Multi-Stage Ranking Exists

Suppose we have:

10,000,000 documents

An advanced relevance model takes:

50 milliseconds

to deeply evaluate one query–document pair.

Running it against all documents would be impractical.

Instead:

10,000,000 Documents
↓
Fast Retriever
↓
500 Candidates
↓
More Expensive Ranker
↓
50 Candidates
↓
Even Stronger Reranker
↓
Top 10

This is a multi-stage ranking pipeline.

Early Stages
→ cheap and broad

Later Stages
→ expensive and precise

9. What Is Reranking?

Reranking is the process of taking an existing candidate set or ranked list and evaluating it again using a stronger or different relevance model.

Initial Ranked Candidates
↓
Reranker
↓
Improved Ranking

For example:

Initial retrieval:

1. A
2. B
3. C
4. D

After reranking:

1. C
2. A
3. B
4. D

The candidates may remain the same.

What changes is:

their order

10. Why Reranking Can Be Better

Initial retrieval uses relatively limited evidence.

For example:

BM25
→ term statistics

or:

Dense Retrieval
→ precomputed embedding similarity

A reranker can examine the relationship between the query and candidate in more detail.

Instead of asking only:

Are these embeddings close?

it can ask something closer to:

How well does this exact document satisfy this exact query?

That deeper comparison can improve precision.

11. Bi-Encoder Review

Chapter 9 introduced the bi-encoder.

A bi-encoder processes query and document separately:

Query
↓
Encoder
↓
Query Vector

and:

Document
↓
Encoder
↓
Document Vector

Then:

Query Vector
↔
Document Vector
↓
Similarity Score

Its main advantage is efficiency.

Document vectors can be precomputed.

This makes bi-encoders excellent for:

large-scale candidate generation

But separate encoding creates a limitation.

12. The Limitation of Separate Encoding

When a document is embedded in advance, the embedding model does not know which future query will be asked.

It creates one general-purpose document representation.

Document
↓
One Dense Vector

That vector may need to represent many aspects of the document.

Then different queries compare against the same representation.

This is efficient.

But it limits how deeply the model can analyze:

this specific query
against
this specific document

This motivates another architecture:

the cross-encoder.

13. What Is a Cross-Encoder?

A cross-encoder processes the query and candidate document together.

Instead of:

Query → Vector A
Document → Vector B
Compare A and B

we do:

Query
+
Document
↓
Model
↓
Relevance Score

For example:

Query:
How can I stop automatic renewal?

Candidate:
To disable auto-renew, open Billing Settings...

The model receives both together.

It can directly examine their relationship.

14. Bi-Encoder vs. Cross-Encoder

Bi-Encoder

Query
↓
Vector

Document
↓
Vector

↓
Similarity

Advantages:

fast
scalable
document embeddings can be precomputed

Best suited for:

candidate generation

Cross-Encoder

Query + Document
↓
Joint Processing
↓
Relevance Score

Advantages:

deeper query–document interaction
more precise relevance estimation

But:

more computationally expensive

Best suited for:

reranking smaller candidate sets

15. Why Cross-Encoders Are More Expensive

Suppose we have:

1,000,000 documents

With a bi-encoder:

Document Embeddings
→ already stored

New Query
→ embed once

Then
→ vector search

With a cross-encoder:

Query + Document 1
→ model

Query + Document 2
→ model

Query + Document 3
→ model

...

Each query–document pair requires processing.

So running a cross-encoder over millions of documents is usually expensive.

But running it over:

Top 50
Top 100
Top 500

candidates can be practical.

16. Joint Query–Document Interaction

Why can cross-encoders rank more precisely?

Because query and document tokens can interact during model processing.

Query Tokens
↕
Document Tokens
↓
Joint Contextual Processing
↓
Relevance Judgment

Suppose the query is:

Apple repair in Toronto

Candidate A:

Apple iPhone repair services in Toronto.

Candidate B:

How to repair apple trees in Ontario.

A stronger joint model can use context to better distinguish:

Apple company

from:

apple fruit/tree

17. Reranking Example

Suppose hybrid retrieval returns:

1. Apple repair guide
2. Toronto phone repair shop
3. Apple tree maintenance
4. iPhone and Mac repair Toronto

The initial ranking may be noisy.

A cross-encoder reranker examines each candidate with:

Query:
Apple repair Toronto

and might produce:

iPhone and Mac repair Toronto → 0.96
Toronto phone repair shop     → 0.81
Apple repair guide            → 0.63
Apple tree maintenance        → 0.08

New ranking:

1. iPhone and Mac repair Toronto
2. Toronto phone repair shop
3. Apple repair guide
4. Apple tree maintenance

The reranker corrected the initial ordering.

18. Reranking Does Not Usually Search the Whole Corpus

This distinction matters.

A reranker normally receives:

existing candidates

It does not usually perform:

full corpus retrieval
Retriever
→ finds candidates

Reranker
→ reevaluates candidates

If the correct document never reaches the candidate set:

Reranker
cannot rescue it

This is why retrieval recall remains important.

19. Retrieval Recall and Ranking Precision

This creates a useful division of responsibility.

Early retrieval should avoid missing useful documents.

So it often emphasizes:

recall

Later reranking tries to put the best documents at the top.

So it often emphasizes:

precision at the top
Retriever
→ Don't miss good candidates

Reranker
→ Put the best candidates first

This is one of the most useful mental models for multi-stage retrieval.

20. Ranking Can Use More Than Text Similarity

A ranking system does not have to use only:

BM25

or:

embedding similarity

It may combine many signals.

Examples:

lexical relevance
semantic relevance
freshness
document quality
authority
location
language
popularity
personalization
structured metadata

The exact signals depend on the application.

For example, local search might strongly care about:

distance
location
business category

while news search may care heavily about:

freshness

21. Ranking Features

A ranking feature is an input signal used to help determine ranking.

Suppose the query is:

Persian dentist Toronto

Possible features could include:

BM25 score
dense similarity
city match
category match
document freshness
business rating
distance
Candidate
↓
Multiple Features
↓
Ranking Model
↓
Score

The system can learn or define how much each feature matters.

22. Retrieval Score vs. Ranking Features

A retrieval score can itself become one ranking feature.

For example:

BM25 Score
Dense Score
Freshness Score
Authority Score
↓
Final Ranker

So BM25 does not need to disappear after candidate generation.

Its score can continue to contribute later.

The same applies to dense similarity.

23. What Is Learning to Rank?

Learning to Rank, often abbreviated LTR, refers to machine-learning approaches that learn how to order results using training data.

Instead of manually specifying:

BM25 × 0.4
Freshness × 0.2
Authority × 0.4

a model can learn useful relationships from examples.

Queries
+
Candidate Documents
+
Relevance Labels / User Signals
↓
Training
↓
Ranking Model

Then:

New Query
+
Candidates
↓
Ranking Model
↓
Ordered Results

24. Where Do Ranking Labels Come From?

Training data may contain relevance judgments such as:

Query
Document
Relevance

For example:

Query:
Persian dentist Toronto

Document A → Highly Relevant
Document B → Somewhat Relevant
Document C → Irrelevant

Labels may come from sources such as:

human relevance judgments
expert annotations
historical user interactions
click data
task-specific feedback

But user behavior data can contain bias and noise.

For example:

clicked
≠
necessarily most relevant

So training signals must be interpreted carefully.

25. Pointwise Ranking

One conceptual approach is pointwise ranking.

The model evaluates each query–document pair independently.

Query + Document A
↓
Score 0.91

Query + Document B
↓
Score 0.63

Query + Document C
↓
Score 0.20

Then results are sorted by score.

The model asks something like:

How relevant is this candidate?

26. Pairwise Ranking

In pairwise ranking, the model focuses on comparing two candidates.

For this query:

Is A better than B?

Then:

A > B

or:

B > A

Training can use many such comparisons to learn a useful ranking function.

The focus is:

relative preference

rather than an isolated score.

27. Listwise Ranking

In listwise ranking, the ranking problem is treated more directly as ordering an entire group of candidates.

Query
+
[A, B, C, D]
↓
Ranking Model
↓
[C, A, D, B]

The objective concerns the quality of the result list itself.

You do not need the mathematical details here.

The important distinction is:

Pointwise
→ score each item

Pairwise
→ compare items

Listwise
→ optimize the ordered list

28. Cross-Encoder Reranking vs. Learning to Rank

These concepts can overlap but are not identical.

A cross-encoder describes a model architecture:

Query + Document
↓
Joint Model
↓
Score

Learning to Rank describes a broader training approach for ordering results.

A ranking system could use:

traditional LTR
neural rerankers
cross-encoders
specialized ranking models

depending on the application.

Cross-Encoder
≠
Learning to Rank

although a cross-encoder can be trained for ranking.

29. Reranking After Hybrid Retrieval

Now connect Chapter 10 directly to Chapter 11.

Hybrid retrieval:

BM25
+
Dense Retrieval
↓
Fusion
↓
200 Candidates

Then:

200 Candidates
↓
Cross-Encoder / Reranker
↓
Top 20

This is a common conceptual architecture:

Broad Retrieval
↓
Strong Candidate Coverage
↓
Precise Reranking
↓
High-Quality Top Results

30. Why Top Results Matter So Much

Suppose the correct document is ranked:

#2

versus:

#87

Technically, the retrieval system found it in both cases.

But practically, those outcomes are very different.

Users may inspect only:

top few results

and downstream AI systems may only pass:

top 5
top 10
top 20

documents into later stages.

Being Retrieved
≠
Being Highly Ranked

This distinction becomes important for AI Visibility.

31. Ranking and the Context Window

Suppose a RAG system retrieves:

100 passages

but can only place:

8 passages

into the LLM context.

Then ranking determines:

which 8 survive
100 Candidates
↓
Ranking / Reranking
↓
Top 8
↓
LLM Context

So ranking can directly influence which evidence the LLM sees.

32. Ranking Errors Can Become Generation Errors

Suppose:

Passage A
→ correct and highly relevant

Passage B
→ partially relevant but outdated

If ranking produces:

B > A

and only B enters the context, the LLM may generate a worse answer.

Ranking Error
↓
Wrong Evidence Selected
↓
Possible Generation Error

The LLM cannot use evidence it never receives.

33. Relevance Is Not the Only Ranking Objective

Depending on the system, final ranking may optimize for more than pure topical relevance.

A system may also care about:

freshness
authority
diversity
geographic suitability
safety
personalization
source quality

For example, consider:

latest tax rules Netherlands

A highly relevant article from:

2019

may be less useful than an equally relevant document from:

2026

Freshness therefore becomes important.

34. Relevance vs. Authority

Suppose two pages answer:

What documents are required for a visa application?

Page A:

random forum post

Page B:

official government documentation

Both may be highly relevant.

But they do not have equal authority.

Relevance
≠
Authority

A ranking system may use authority as an additional signal.

35. Relevance vs. Freshness

Highly Relevant
≠
Current

A page may perfectly discuss:

subscription pricing

but contain pricing from three years ago.

A modern search system may need to consider:

relevance
+
freshness

especially for time-sensitive queries.

36. Ranking vs. Factual Verification

Ranking answers:

Which candidate appears most useful for this query?

Factual verification asks:

Is the information actually correct?

Those are different problems.

A false document can still be:

highly relevant
well-written
semantically similar
lexically precise
Rank 1
≠
Guaranteed Truth

This distinction will matter throughout the rest of the course.

37. Diversity in Ranking

Sometimes the top results can be too similar.

Suppose a search system returns:

1. Article A
2. Copy of Article A
3. Syndicated Article A
4. Another version of Article A
5. Article B

Even if all are relevant, the result set may provide little diversity.

Some systems therefore try to balance:

relevance
+
diversity
Top Results
should not always mean
Near-Duplicates of the Same Information

This can be useful in search and evidence gathering.

38. Reranking Can Be Query-Aware in Greater Detail

Initial retrieval often uses compact representations.

A reranker can analyze details such as:

negation
exact intent
specific relationships
entity roles
question type
important constraints

For example:

Query:
Which plans do not include API access?

Candidate A:

The Starter plan includes API access.

Candidate B:

The Free plan does not include API access.

Both discuss:

plan
API access

But Candidate B directly satisfies the query.

A strong reranker can better capture that distinction.

39. Reranking Is Still Not Reasoning Over Everything

A reranker can be more sophisticated than the retriever.

But it still has limits.

For example, it may not fully resolve:

conflicting sources
complex multi-document reasoning
factual disputes
source credibility
temporal inconsistencies
Better Reranking
≠
Complete Answer Verification

Its main task remains:

better candidate ordering

40. LLM-Based Reranking

Modern systems can also use an LLM or LLM-like model to evaluate candidate relevance.

Query
+
Candidate
↓
LLM-Based Relevance Evaluation
↓
Score / Preference

Or:

Query
+
Several Candidates
↓
LLM
↓
Ordered Candidates

This can capture complex instructions and relevance relationships.

But it can also be:

slower
more expensive
less deterministic

than simpler rankers.

So LLM reranking is another option, not automatically the best choice for every system.

41. A Complete Simplified Ranking Pipeline

Suppose a support system contains:

5,000,000 passages

The user asks:

How can I cancel automatic renewal on my yearly subscription?

Step 1 — Candidate Generation

Hybrid retrieval runs:

BM25
+
Dense Retrieval
↓
500 Candidates

Step 2 — Fusion

RRF or another fusion method combines the retrieval lists.

500 Candidates
↓
Hybrid Ordering

Step 3 — First Ranking Cut

The system keeps:

Top 100

Step 4 — Cross-Encoder Reranking

For each candidate:

Query + Passage
↓
Cross-Encoder
↓
Relevance Score

Example:

Auto-Renew Settings       → 0.97
Annual Subscription FAQ   → 0.88
Refund Policy             → 0.61
Payment Failure Guide     → 0.24

Step 5 — Final Ranking

1. Auto-Renew Settings
2. Annual Subscription FAQ
3. Refund Policy
4. Payment Failure Guide

Step 6 — Top Results Continue

If this is a search engine:

Top Results
→ shown to user

If this is RAG:

Top Passages
→ inserted into LLM context

The full pipeline is:

Corpus
↓
Candidate Generation
↓
Hybrid Retrieval
↓
Initial Ranking
↓
Reranking
↓
Top Evidence
↓
Search Results or LLM Context

43. Ranking and AI Visibility

Now we can make an important connection to AI Visibility.

A page may be:

crawlable
indexed
retrievable

and still never influence an AI answer.

Why?

Because it may repeatedly lose during ranking.

Page Exists
↓
Indexed
↓
Retrieved as Candidate
↓
Ranked Too Low
↓
Not Selected as Evidence
↓
No Influence on Final Answer

So AI Visibility is not only about:

being present in the corpus

or:

being retrievable

It can also depend on:

winning later ranking stages

We will develop this idea much further later in the course.

44. Common Misunderstandings

Misunderstanding 1: Retrieval and ranking are the same thing

Not exactly. Retrieval finds candidates; ranking determines their order.

Misunderstanding 2: Retrieval results have no ranking

Incorrect. BM25 and dense retrieval often produce an initial ranking themselves.

Misunderstanding 3: Reranking means searching the corpus again

Usually not. Reranking typically evaluates candidates already retrieved.

Misunderstanding 4: A reranker can recover a document that was never retrieved

Incorrect. If the document never reaches the candidate set, the reranker usually cannot evaluate it.

Misunderstanding 5: Cross-encoders are ideal for searching millions of documents directly

Usually not. Their deeper joint processing makes them better suited to smaller candidate sets.

Misunderstanding 6: Bi-encoders are worse models

Incorrect. They solve a different problem: efficient large-scale retrieval.

Misunderstanding 7: Cross-encoder and Learning to Rank mean the same thing

Incorrect. One describes a model architecture; the other is a broader ranking methodology.

Misunderstanding 8: Ranking only considers textual similarity

Incorrect. Ranking can use many features, including freshness, authority, metadata, location, and other signals.

Misunderstanding 9: Rank 1 means the information is true

Incorrect. Ranking estimates usefulness or relevance, not factual truth.

Misunderstanding 10: More complex reranking always improves the system

Incorrect. Quality must be evaluated against latency, cost, and actual retrieval performance.

Misunderstanding 11: If a document was retrieved, its job is done

Incorrect. A retrieved document may rank too low to ever reach the user or LLM context.

45. How the Concepts Connect

Chapter 7 gave us:

Corpus
↓
Candidate Generation
↓
Ranking
↓
Results

Chapter 8 showed one candidate-generation method:

Lexical Retrieval
↓
BM25

Chapter 9 added:

Dense Retrieval
↓
Vector Similarity

Chapter 10 combined them:

Lexical
+
Dense
↓
Hybrid Retrieval
↓
Candidate Set

Chapter 11 now adds deeper ordering:

Hybrid Candidate Set
↓
Initial Ranking
↓
Reranking
↓
Top Results

A common modern retrieval architecture can therefore be represented as:

Corpus
↓
Indexes
↓
User Query
↓
Lexical + Dense Retrieval
↓
Fusion
↓
Candidate Set
↓
Cross-Encoder / Ranking Model
↓
Reranking
↓
Top-K Results

The next major question is:

How do we know whether this retrieval and ranking system is actually good?

That leads directly into Chapter 12:

Retrieval Evaluation and Ranking Metrics.

Mental Model

Imagine a hiring process.

A company receives:

100,000 applications

It would be too expensive for senior interviewers to deeply evaluate all of them.

So the first stage performs broad filtering:

100,000 Applications
↓
Fast Screening
↓
1,000 Candidates

This is similar to:

retrieval

Then a stronger screening process reduces them further:

1,000
↓
Detailed Review
↓
100

Then interviews produce:

100
↓
Deep Evaluation
↓
Top 10

This is similar to:

reranking

The early system needs to avoid throwing away strong candidates.

The later system can spend more effort evaluating the survivors.

Retriever
→ broad search

Ranker
→ ordering

Reranker
→ deeper ordering of fewer candidates

The same principle applies to search.

You cannot afford to deeply analyze every document in the corpus.

Instead:

Huge Corpus
↓
Cheap Broad Retrieval
↓
Small Candidate Set
↓
Expensive Precise Evaluation
↓
Best Results

And the simplest reliable definition is:

Ranking determines the order of retrieved candidates, while reranking applies a stronger and usually more expensive evaluation to a smaller candidate set in order to improve the final ordering.

Chapter 11 Recap

Ranking

Ordering candidate results according to estimated relevance or usefulness.

Candidates
↓
Scores
↓
Ordered Results
Ranker

A component or model that evaluates candidates and produces ranking scores or an ordering.

Ranking Score

A numerical signal used to order results.

Higher Score
→ usually higher rank

within that ranking system.

Retrieval vs. Ranking
Retrieval
→ What should we consider?

Ranking
→ Which candidates should come first?
First-Stage Ranking

The initial ordering produced by the retrieval system.

Examples:

BM25 ranking
dense similarity ranking
hybrid fusion ranking
Multi-Stage Ranking

Using progressively stronger models on progressively smaller candidate sets.

Large Corpus
↓
Fast Retriever
↓
Candidates
↓
Stronger Ranker
↓
Smaller Set
↓
Reranker
↓
Top Results
Reranking

Reevaluating an existing candidate set using a stronger or different relevance model.

Initial Ranking
↓
Reranker
↓
Improved Ranking
Bi-Encoder

Processes query and document separately.

Query → Query Vector

Document → Document Vector

↓
Similarity

Strength:

efficient large-scale retrieval
Cross-Encoder

Processes query and document together.

Query + Document
↓
Joint Processing
↓
Relevance Score

Strength:

more precise query–document evaluation

Cost:

more computation
Retriever vs. Reranker
Retriever
→ broad candidate generation

Reranker
→ deeper evaluation of retrieved candidates
Ranking Features

Signals used by a ranking system.

Examples:

BM25
dense similarity
freshness
authority
location
metadata
Learning to Rank — LTR

Machine-learning methods that learn how to order results from relevance data or behavioral signals.

Training Data
↓
Learn Ranking Function
↓
Rank New Candidates
Pointwise Ranking
Query + Candidate
↓
Individual Score
Pairwise Ranking
A vs. B
↓
Which is better?
Listwise Ranking
[A, B, C, D]
↓
Optimize Entire Ordered List
Retrieval Recall vs. Ranking Precision

A useful division of responsibilities:

Retriever
→ avoid missing relevant information

Reranker
→ put the strongest information at the top
Ranking and RAG
Retrieved Passages
↓
Ranking / Reranking
↓
Top Passages
↓
LLM Context

The LLM only has access to evidence that survives this selection.

Ranking Is Not Truth Verification
High Rank
≠
Guaranteed Truth

A ranking system estimates usefulness or relevance.

The complete Chapter 11 pipeline is:

Corpus
↓
Lexical / Dense Indexes
↓
User Query
↓
Candidate Generation
↓
Lexical + Dense Retrieval
↓
Fusion
↓
Initial Ranked Candidates
↓
Stronger Ranker / Cross-Encoder
↓
Reranking
↓
Top-K Results
↓
Search Results or LLM Context

We now know how information moves from:

millions of possible documents

to:

a small ordered set of strong candidates

The next question is:

How do we measure whether retrieval and ranking are actually working well?

That leads into Chapter 12:

Retrieval Evaluation and Ranking Metrics.

The Single Most Important Idea from Chapter 11

Retrieval finds a manageable set of candidates, ranking orders those candidates, and reranking uses a stronger—usually more expensive—model to evaluate a smaller set more precisely. This multi-stage design allows modern search and RAG systems to balance scalability with high-quality relevance at the top of the result list.

← Back to AI Visibility Guide