Chapter 10- Hybrid Retrieval

Chapter Goal

In Chapter 8, we learned lexical retrieval.

Lexical retrieval is strong when exact words matter:

Query Terms
↓
Inverted Index
↓
BM25 / Lexical Scoring
↓
Ranked Results

It performs especially well for:

names
codes
identifiers
exact phrases
rare terms
precise terminology

But it can struggle when the query and relevant document use different vocabulary.

For example:

Query:
cheap private teacher

Document:
affordable one-to-one tutor

In Chapter 9, we learned dense retrieval.

Dense retrieval represents queries and documents as embeddings:

Query
↓
Embedding
↓
Query Vector

Documents
↓
Embeddings
↓
Vector Index

Then it retrieves semantically similar documents.

This helps with:

synonyms
paraphrases
different wording
conceptual similarity

But dense retrieval can struggle with precise distinctions such as:

HR-204
Model 15 vs. Model 16
Toronto vs. Vancouver
specific error codes
exact product names

So we now have two systems with complementary strengths.

That leads to a natural question:

Why choose only one?

A retrieval system can use both.

This is called Hybrid Retrieval.

At the simplest level:

Lexical Retrieval
+
Dense Retrieval
↓
Combine Results
↓
Better Candidate Set

Hybrid retrieval attempts to preserve:

exact lexical matching

while also capturing:

semantic similarity

By the end of this chapter, you should understand:

what hybrid retrieval is;
why lexical and dense retrieval complement each other;
how parallel hybrid retrieval works;
what candidate union means;
what candidate intersection means;
why retrieval scores cannot always be directly added;
what score normalization means;
what weighted score fusion is;
what rank fusion is;
what Reciprocal Rank Fusion is;
what RRF means;
how hybrid retrieval can improve recall;
why more candidates do not automatically mean better results;
how hybrid retrieval relates to ranking and reranking;
why metadata filters can still matter;
and where hybrid retrieval fits inside modern search and RAG systems.

The central pipeline of this chapter is:

User Query
├──→ Lexical Retriever
│     ↓
│     Lexical Candidates
│
└──→ Dense Retriever
      ↓
      Semantic Candidates

↓
Fusion
↓
Combined Candidate Set
↓
Ranking / Reranking
↓
Final Results

1. Why Hybrid Retrieval Exists

Consider the query:

affordable immigration lawyer Toronto

Suppose Document A says:

Affordable immigration lawyer in Toronto.

Lexical retrieval should find it easily.

Now Document B says:

Low-cost Toronto attorney specializing in visa and permanent residence cases.

The meaning is highly relevant, but the wording differs:

affordable
≈
low-cost

lawyer
≈
attorney

immigration
≈
visa / permanent residence

Dense retrieval may find Document B more easily.

Lexical Retriever
→ strong for A

Dense Retriever
→ strong for B

If we use both:

A + B

can enter the candidate set.

That is the basic motivation for hybrid retrieval.

2. What Is Hybrid Retrieval?

Hybrid retrieval combines two or more retrieval approaches to produce a stronger set of candidate results.

In this chapter, we focus mainly on combining:

Lexical Retrieval
+
Dense Retrieval

Query
↓
Two Retrieval Paths
↓
Two Result Sets
↓
Combine
↓
Hybrid Results

The lexical retriever looks for:

word-level evidence

The dense retriever looks for:

embedding-level similarity

Hybrid retrieval uses both signals.

3. Hybrid Retrieval Is Not a New Representation Type

Do not think of hybrid retrieval as a third type of vector.

It is not:

Sparse Vector
Dense Vector
Hybrid Vector

Instead, hybrid retrieval is a retrieval strategy.

It combines outputs from multiple retrieval systems.

BM25 Retriever
+
Dense Vector Retriever
↓
Fusion

Hybrid Retrieval
≠
new embedding type

It is a method for combining evidence.

4. The Two Signals

A simple way to understand hybrid retrieval is to think in terms of two signals.

Lexical Signal

Do the query and document share important words?

Semantic Signal

Do the query and document express related meaning?

A hybrid system can use both:

Exact Word Evidence
+
Meaning Evidence
↓
Retrieval Decision

This is especially useful because these signals fail in different ways.

5. Complementary Failure Modes

Suppose the query is:

cheap heart doctor Toronto

Document A:

Affordable cardiologist in Toronto.

Dense retrieval may recognize:

cheap
≈
affordable

heart doctor
≈
cardiologist

Lexical retrieval may have weaker overlap.

Now consider:

Cardiology policy HR-204 Toronto

The exact identifier:

HR-204

may be critical.

Lexical retrieval can strongly preserve this exact match.

Dense retrieval may focus too much on the broader cardiology-related meaning.

Lexical Weakness
→ vocabulary mismatch

Dense Weakness
→ exact-detail mismatch

Hybrid retrieval tries to reduce both.

6. Parallel Hybrid Retrieval

One common design runs lexical and dense retrieval independently.

                 ┌→ BM25 Retriever
User Query ──────┤
                 └→ Dense Retriever

The lexical retriever might return:

A
B
C
D

The dense retriever might return:

B
E
F
A

Now the system needs to combine those results.

This is called fusion.

7. What Is Fusion?

Fusion means combining multiple retrieval result sets or ranking signals.

Suppose:

Lexical Results:
A
B
C
D

Dense Results:
B
E
F
A

The fused candidate set might become:

A
B
C
D
E
F

But simply collecting documents is only the first step.

The system must still decide:

Which results should rank highest?

That requires a fusion strategy.

8. Candidate Union

One simple approach is to take the union of both result sets.

Suppose:

Lexical:
A, B, C

Dense:
B, D, E

Union:

A, B, C, D, E

Union
=
all unique candidates from both systems

This is useful when the goal is to avoid missing potentially relevant documents.

9. Why Candidate Union Can Improve Recall

Recall asks, conceptually:

Did we retrieve the relevant information?

Suppose the lexical retriever finds:

A
B
C

but misses relevant Document D.

The dense retriever finds:

B
D
E

Combining them gives:

A
B
C
D
E

Now D survives.

Different Retrievers
↓
Different Relevant Documents
↓
Union
↓
Broader Candidate Coverage

10. Candidate Intersection

Another possible operation is intersection.

Suppose:

Lexical:
A, B, C

Dense:
B, C, D

Intersection:

B, C

Intersection
=
candidates supported by multiple retrievers

This can provide strong confidence that the documents have both lexical and semantic support.

11. Why Intersection Can Be Too Strict

Suppose Document D is highly relevant semantically but uses completely different wording.

Dense retrieval finds it.

Lexical retrieval does not.

If we keep only the intersection:

D disappears

Likewise, an exact identifier result found only by BM25 could disappear.

Intersection
→ stronger agreement
but
→ lower candidate coverage

For many retrieval pipelines, union is more useful during candidate generation.

The final ranking can later decide which candidates are strongest.

12. Candidate Generation vs. Fusion Ranking

It helps to separate two questions.

First:

Which candidates should survive?

Second:

In what order should they appear?

Hybrid retrieval may use:

Candidate Union
↓
larger candidate set

and then:

Fusion Scoring / Reranking
↓
better ordering

These are distinct stages.

13. The Score Problem

Suppose BM25 returns:

Document A → 12.8
Document B → 8.4

Dense retrieval returns cosine similarities:

Document A → 0.72
Document B → 0.91

Can we simply add:

12.8 + 0.72

and compare it with:

8.4 + 0.91

Usually, that would be a poor assumption.

Why?

Because the scores live on different scales.

BM25 score
and
cosine similarity

do not have the same numerical meaning.

14. Retrieval Scores Are Not Directly Comparable

Recall:

BM25 = ranking score

and:

Cosine Similarity = vector similarity score

Their values can have completely different distributions.

For example:

BM25:
0 → 25

Dense:
0.4 → 0.95

So directly combining raw scores may cause one system to dominate simply because its numerical range is larger.

This leads to score normalization.

15. Score Normalization

Score normalization transforms scores so that signals from different systems can be combined more meaningfully.

BM25 Scores
↓
Normalize
↓
Comparable Range

Dense Scores
↓
Normalize
↓
Comparable Range

For example:

BM25 Raw:
15.2
9.1
4.3

Normalized:
1.00
0.55
0.00

Dense scores might similarly be transformed.

Then the normalized values can be combined.

Normalization Does Not Make the Scores Identical

Normalization does not mean:

BM25 meaning
=
Dense similarity meaning

It only helps bring their numerical scales into a form that can be combined.

BM25
→ lexical evidence

Dense Score
→ semantic evidence

Normalization helps fusion.

It does not erase the conceptual difference.

16. Weighted Score Fusion

One simple fusion strategy is a weighted combination.

Hybrid Score
=
Lexical Weight × Lexical Score
+
Dense Weight × Dense Score

For example:

Lexical Weight = 0.4
Dense Weight   = 0.6

Then:

Hybrid Score
=
0.4 × normalized lexical score
+
0.6 × normalized dense score

The weights determine how much each retrieval signal contributes.

17. Why Use Different Weights?

Different applications may care about different types of matches.

For example, a legal search system may care strongly about:

case numbers
statute numbers
exact names
citations

So lexical weight may be relatively important.

A support knowledge base may contain many paraphrased user questions:

How do I stop my plan?
How can I end my subscription?
How do I cancel my membership?

Semantic retrieval may deserve more weight.

Optimal Lexical / Dense Balance
depends on the retrieval task

18. Query-Dependent Weighting

The same system does not necessarily need the same weights for every query.

Consider:

Query A:
HR-204

Exact lexical identity is extremely important.

Now:

Query B:
How do I stop paying for my yearly plan?

Semantic retrieval may be more valuable.

A sophisticated system may therefore adapt retrieval behavior based on query characteristics.

Query Type
↓
Choose Retrieval Balance

This is one way query understanding can influence retrieval.

19. Score Fusion vs. Rank Fusion

There are two broad ways to combine retrieval systems.

Score Fusion

Combine actual numerical scores.

Lexical Score
+
Dense Score
↓
Hybrid Score

Rank Fusion

Ignore or reduce dependence on raw score scales and combine the rank positions from different systems.

Lexical Rank
+
Dense Rank
↓
Combined Rank

Rank fusion becomes useful when raw scores are difficult to compare.

20. Why Rank Fusion Can Be Useful

Suppose:

Lexical ranking:

1. A
2. B
3. C

and:

Dense ranking:

1. B
2. D
3. A

Even if BM25 and dense scores have unrelated scales, their rank positions still tell us something.

For example:

A
→ lexical rank 1
→ dense rank 3

B
→ lexical rank 2
→ dense rank 1

Both A and B receive strong support from both systems.

Rank fusion can use that information without needing direct score comparability.

21. Reciprocal Rank Fusion

A common rank-fusion method is:

Reciprocal Rank Fusion, or RRF.

The core idea is simple:

Documents receive more credit when they appear near the top of one or more ranked lists.

High Rank
→ Strong Contribution

Low Rank
→ Smaller Contribution

If a document ranks highly in both lexical and dense retrieval, its combined RRF score can become strong.

22. RRF Intuition

Suppose:

Lexical Ranking:
1. A
2. B
3. C

Dense Ranking:
1. B
2. D
3. A

Document A:

Lexical → Rank 1
Dense   → Rank 3

Document B:

Lexical → Rank 2
Dense   → Rank 1

Document C:

Lexical → Rank 3
Dense   → not highly ranked

Document D:

Dense   → Rank 2
Lexical → not highly ranked

RRF rewards A and B because they perform strongly across multiple ranked lists.

A possible fused ranking might be:

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

The exact result depends on the RRF formula and list positions.

23. RRF Does Not Need Raw Score Normalization

One major advantage of RRF is that it uses rank positions rather than trying to directly combine:

BM25 = 13.7

with:

Cosine = 0.86

So it avoids much of the problem of incompatible score scales.

Raw Scores
→ different scales
→ difficult to combine

Ranks
→ shared ordering concept
→ easier to fuse

This is one reason RRF is widely useful in hybrid retrieval.

24. RRF Does Not Mean Every Retriever Has Equal Quality

Rank fusion can combine multiple retrievers.

But combining weak retrieval systems does not magically produce perfect results.

Poor Lexical Candidates
+
Poor Dense Candidates
↓
Still Poor Candidate Pool

Fusion improves the use of complementary signals.

It cannot recover information absent from all retrieval paths.

25. Hybrid Retrieval and Recall

One major benefit of hybrid retrieval is improved candidate coverage.

Suppose the relevant documents are:

A
B
C
D

Lexical retrieval finds:

A
B

Dense retrieval finds:

B
C
D

The union finds:

A
B
C
D

This can significantly improve recall.

Different Retrieval Signals
↓
Different Relevant Candidates
↓
Combined Candidate Pool
↓
Higher Chance of Capturing Relevant Information

26. Higher Recall Can Increase Noise

More candidates are not automatically better.

Suppose lexical retrieval returns 100 candidates.

Dense retrieval returns another 100.

Their union contains:

170 unique candidates

Now the system may have more relevant information.

But it may also have more:

weakly relevant
duplicate
ambiguous
off-intent

documents.

Higher Recall
can also mean
More Noise

This is why later ranking and reranking stages matter.

27. Hybrid Retrieval Does Not Replace Ranking

Hybrid retrieval mainly improves the candidate-generation stage.

It does not automatically solve final ordering.

A typical pipeline can be:

Lexical Retriever
+
Dense Retriever
↓
Hybrid Candidate Set
↓
Fusion
↓
Reranker
↓
Final Results

Hybrid Retrieval
≠
Final Ranking System

It can provide better candidates for a stronger downstream ranker.

28. Hybrid Retrieval and Reranking

Suppose:

BM25
→ 100 candidates

Dense Retriever
→ 100 candidates

After union:

160 unique candidates

A reranker can then evaluate these candidates more deeply.

Hybrid Candidate Generation
↓
160 Candidates
↓
Reranker
↓
Top 20

This architecture is powerful because:

retrieval
→ broad and fast

reranking
→ narrower and more precise

Chapter 11 will explore this much more deeply.

29. Hybrid Retrieval Can Work at Passage Level

Hybrid retrieval does not need to operate only on complete documents.

Suppose a knowledge base is split into passages.

Each passage may be available through:

Inverted Index

and:

Vector Index

Then:

User Query
├──→ BM25 Passage Retrieval
└──→ Dense Passage Retrieval
↓
Fusion
↓
Relevant Passages

This is especially useful for RAG.

30. Example: Hybrid Retrieval in a Support Knowledge Base

Suppose the user asks:

How can I stop my yearly membership?

The knowledge base contains:

Passage A:
How to cancel an annual subscription

Passage B:
Annual subscription refund policy

Passage C:
How to stop automatic renewal

Passage D:
Monthly membership pricing

Lexical retrieval may favor:

Passage C

because it contains:

stop

Dense retrieval may strongly favor:

Passage A

because:

stop yearly membership
≈
cancel annual subscription

A hybrid system can preserve both.

Then a reranker may determine that Passage A best answers the user's actual request.

31. Example: Hybrid Retrieval with an Identifier

Now suppose the query is:

HR-204 parental leave

Lexical retrieval strongly identifies:

HR-204

Dense retrieval may find additional semantically related passages about:

maternity leave
parental leave
family leave
employee leave policy

The combined candidate pool may contain:

exact HR-204 passage
+
semantically related policy passages

This is a good example of lexical precision and semantic coverage working together.

32. Metadata Still Matters

Hybrid retrieval does not mean every retrieval decision should come only from BM25 and embeddings.

Suppose the user asks:

Persian dentist in Toronto

The system might use:

BM25
+
Dense Retrieval
+
Metadata:
city = Toronto
category = Dentist

Lexical Evidence
+
Semantic Evidence
+
Structured Constraints
↓
Better Candidate Selection

Hybrid retrieval can therefore sit inside a broader retrieval architecture.

33. Filtering Before Retrieval

Sometimes metadata filters are applied before vector or lexical retrieval.

For example:

city = Toronto

first reduces the search space.

Then:

BM25
+
Dense Search

operate only on Toronto documents.

Corpus
↓
Metadata Filter
↓
Smaller Corpus
↓
Hybrid Retrieval

This can improve efficiency and precision.

34. Filtering After Retrieval

Another design retrieves candidates first and applies filters later.

Hybrid Retrieval
↓
Candidate Set
↓
Metadata Filtering

Which design is better depends on:

system architecture
index capabilities
filter selectivity
latency requirements
retrieval quality

The important point is that hybrid retrieval and structured filtering can work together.

35. Hybrid Retrieval Is Not Just “BM25 + Embeddings”

The most common example is:

BM25
+
Dense Embeddings

But hybrid retrieval is a broader idea.

A system could combine multiple retrieval signals such as:

lexical retrieval
dense retrieval
entity matching
metadata filters
geographic constraints
freshness signals
specialized retrievers

In this textbook, however, when we say:

hybrid retrieval

we will usually mean:

lexical
+
dense

unless stated otherwise.

36. Query Routing vs. Hybrid Retrieval

Another related idea is query routing.

Hybrid retrieval says:

Use multiple retrievers
and combine their outputs.

Query routing says:

Inspect the query
↓
Choose which retriever or retrieval strategy to use

For example:

HR-204
→ lexical-heavy retrieval

while:

How can I stop paying for my plan?
→ semantic-heavy retrieval

These strategies can also be combined.

A system might use hybrid retrieval but dynamically adjust the balance.

37. Hybrid Retrieval, Generation, and Truth

Hybrid Retrieval Does Not Guarantee the Correct Answer

Suppose hybrid retrieval finds the correct document.

That does not guarantee that a later LLM will:

interpret it correctly
cite it correctly
avoid unsupported claims
choose the right passage

Better Retrieval
≠
Guaranteed Correct Generation

Hybrid retrieval improves candidate selection.

Generation remains a separate stage.

Hybrid Retrieval Does Not Verify Truth

Suppose both BM25 and dense retrieval strongly rank a false article.

That article may receive strong hybrid support.

Lexical Match
+
Semantic Match
≠
Factual Truth

Retrieval evaluates relevance.

Truth, authority, and evidence quality require additional mechanisms.

38. A Complete Simplified Hybrid Retrieval Example

Suppose a company has:

1,000,000 support passages

The user asks:

How do I stop automatic payment for my yearly plan?

The best passage says:

To disable automatic renewal for an annual subscription, open Billing Settings and turn off Auto-Renew.

Let us follow the process.

Step 1 — Lexical Retrieval

The query contains terms such as:

stop
automatic
payment
yearly
plan

BM25 may retrieve passages mentioning:

automatic payment
yearly plan
billing

Suppose its top results are:

A
B
C
D

Step 2 — Dense Retrieval

The query becomes an embedding.

Dense retrieval connects:

stop automatic payment
≈
disable auto-renew

yearly plan
≈
annual subscription

Its results are:

C
E
A
F

Step 3 — Candidate Fusion

Union:

A
B
C
D
E
F

Step 4 — Fusion Ranking

A rank-fusion or score-fusion method combines evidence from both systems.

Suppose:

A
→ strong lexical
→ strong semantic

C
→ strong lexical
→ strongest semantic

E
→ weak lexical
→ strong semantic

The hybrid ordering might become:

1. C
2. A
3. E
4. B
5. F
6. D

Step 5 — Reranking

A more precise reranker evaluates the top candidates.

It recognizes that:

disable automatic renewal

directly satisfies the user's intent.

Final results:

1. Auto-Renew Settings
2. Annual Subscription Management
3. Billing FAQ

This shows the role of hybrid retrieval clearly:

Lexical Coverage
+
Semantic Coverage
↓
Better Candidate Pool
↓
Reranking
↓
Better Final Ordering

39. When Hybrid Retrieval Helps Most

Hybrid retrieval is especially useful when the corpus contains a mixture of:

natural-language descriptions
technical terms
names
identifiers
codes
synonyms
paraphrases
location names
specialized terminology

For example:

enterprise search
customer support
e-commerce search
legal search
medical search
RAG knowledge bases
AI search

often benefit from multiple retrieval signals.

40. When Hybrid Retrieval May Be Unnecessary

Not every search system needs maximum complexity.

Suppose a small database is searched almost entirely by exact product IDs:

SKU-8821
SKU-9924
SKU-1172

Lexical or structured lookup may already work extremely well.

Likewise, a narrow semantic FAQ system may perform adequately with dense retrieval alone.

More Retrieval Components
≠
Automatically Better System

Architecture should match the problem.

41. Cost and Complexity

Hybrid retrieval adds additional system complexity.

Instead of maintaining only:

Inverted Index

or only:

Vector Index

you may maintain both.

At query time, you may run:

Lexical Retrieval
+
Dense Retrieval
+
Fusion
+
Reranking

This can increase:

latency
compute
engineering complexity
evaluation complexity

So hybrid retrieval is useful when the quality improvement justifies the added complexity.

42. Hybrid Retrieval Must Be Evaluated

You cannot assume that:

BM25 + Dense

will automatically outperform either one alone for every application.

A proper comparison might evaluate:

Lexical Only

Dense Only

Hybrid

on real queries.

Later we will study metrics such as:

Recall
Precision
MRR
NDCG

These help determine whether the hybrid system actually performs better.

43. Common Misunderstandings

Misunderstanding 1: Hybrid retrieval is a new kind of embedding

Incorrect. Hybrid retrieval combines multiple retrieval approaches.

Misunderstanding 2: Hybrid retrieval means every document must match both lexical and dense retrieval

Incorrect. A candidate may come from only one retriever and still survive through union or fusion.

Misunderstanding 3: Intersection is always better because both retrievers agree

Incorrect. Intersection can remove highly relevant documents that only one retriever discovers.

Misunderstanding 4: BM25 and cosine similarity can always be added directly

Incorrect. Their scores have different meanings and scales.

Misunderstanding 5: Score normalization makes lexical and dense signals conceptually identical

Incorrect. Normalization only helps combine their numerical values.

Misunderstanding 6: Weighted fusion always uses equal weights

Incorrect. Weights can depend on the application and even on the query.

Misunderstanding 7: Rank fusion uses raw scores

Not necessarily. Methods such as RRF operate primarily on ranking positions.

Misunderstanding 8: RRF is a reranker

Not exactly. RRF is a method for fusing ranked result lists. A separate reranker may still be applied afterward.

Misunderstanding 9: Hybrid retrieval eliminates retrieval failures

Incorrect. If every retriever misses the relevant document, fusion cannot recover it.

Misunderstanding 10: More candidates always improve retrieval

Incorrect. More candidates can increase recall but also increase noise.

Misunderstanding 11: Hybrid retrieval replaces metadata filtering

Incorrect. Structured filters and hybrid retrieval can work together.

Misunderstanding 12: Hybrid retrieval guarantees factual correctness

Incorrect. Relevance and truth remain separate problems.

Misunderstanding 13: Hybrid retrieval removes the need for reranking

Incorrect. Hybrid retrieval can produce a stronger candidate set, but final ordering may still benefit from reranking.

Misunderstanding 14: Dense retrieval should always receive more weight because it is newer

Incorrect. Exact lexical evidence can be critical for many tasks.

44. How the Concepts Connect

Chapter 7 introduced the general retrieval architecture:

Query
↓
Candidate Generation
↓
Ranking
↓
Results

Chapter 8 introduced lexical retrieval:

Query Terms
↓
Inverted Index
↓
BM25
↓
Lexical Candidates

Chapter 9 introduced dense retrieval:

Query
↓
Embedding Model
↓
Query Vector
↓
Vector Index
↓
Dense Candidates

Chapter 10 combines them:

                    ┌→ Lexical Retrieval
User Query ─────────┤
                    └→ Dense Retrieval

↓
Fusion
↓
Hybrid Candidate Set
↓
Ranking / Reranking
↓
Final Results

The key relationship is:

Lexical Retrieval
→ protects exact lexical evidence

Dense Retrieval
→ captures semantic relationships

Hybrid Retrieval
→ uses both

The next question is:

Once we have a good candidate set, how do we evaluate those candidates more precisely and decide the final order?

That leads directly into Chapter 11:

Ranking and Reranking.

Mental Model

Imagine two librarians searching the same giant library.

The first librarian is obsessed with exact words.

You ask:

HR-204 parental leave

and this librarian immediately finds the documents containing:

HR-204

This is the lexical retriever.

The second librarian focuses more on meaning.

You ask:

How much time can new parents take away from work?

and this librarian finds documents about:

parental leave
maternity leave
family leave

even when the exact words differ.

This is the dense retriever.

Now imagine they both hand their results to a third person.

Lexical librarian:

A
B
C

Semantic librarian:

B
D
E

The third person combines them:

A
B
C
D
E

Then asks:

Which results have the strongest combined evidence?

A document may be strong because:

it matches the exact words

or:

it strongly matches the meaning

or, ideally:

both

That combination is hybrid retrieval.

The simplest mental model is:

Lexical Retriever
→ exact words

Dense Retriever
→ related meaning

Fusion
→ combine evidence

Reranking
→ choose the best final order

And the simplest reliable definition is:

Hybrid retrieval combines lexical and semantic retrieval so that the system can preserve exact term matches while also finding relevant information expressed with different wording. The resulting candidates are then fused and can be ranked or reranked before final use.

Chapter 10 Recap

Hybrid Retrieval

A retrieval strategy that combines multiple retrieval approaches.

Lexical Retrieval
+
Dense Retrieval
↓
Hybrid Retrieval

Lexical Signal

Evidence based primarily on shared words and terms.

Query Terms
↔
Document Terms

Semantic Signal

Evidence based on similarity between learned representations.

Query Embedding
↔
Document Embedding

Parallel Hybrid Retrieval

Running lexical and dense retrieval independently for the same query.

Query
├──→ BM25
└──→ Dense Retriever

Fusion

The process of combining outputs from multiple retrievers.

Lexical Results
+
Dense Results
↓
Fusion

Candidate Union

Keeping all unique candidates returned by either retriever.

A, B, C
+
B, D, E
↓
A, B, C, D, E

Union can improve candidate coverage.

Candidate Intersection

Keeping only candidates returned by both retrievers.

A, B, C
∩
B, C, D
↓
B, C

Intersection can be more restrictive and may remove useful candidates.

Score Fusion

Combining numerical retrieval scores.

Lexical Score
+
Dense Score
↓
Hybrid Score

Score Normalization

Transforming different score ranges before combining them.

BM25 Scores
↓
Normalization

Dense Scores
↓
Normalization

↓
More Comparable Signals

Weighted Fusion

Assigning different importance to lexical and dense evidence.

Hybrid Score
=
w1 × Lexical Score
+
w2 × Dense Score

Rank Fusion

Combining result lists based on rank positions rather than raw score values.

Lexical Rank
+
Dense Rank
↓
Combined Rank

Reciprocal Rank Fusion — RRF

A rank-fusion method that rewards documents appearing near the top of one or more ranked lists.

High Rank
→ Stronger Contribution

RRF is useful because it does not require BM25 scores and dense similarity scores to use the same numerical scale.

Recall Benefit

Hybrid retrieval can retrieve documents missed by one individual retriever.

Lexical Candidates
+
Dense Candidates
↓
Broader Coverage

Recall vs. Noise

More candidates may improve recall but also introduce additional weak results.

More Coverage
↔
More Noise

This is why ranking and reranking remain important.

Hybrid Retrieval and Reranking

A common architecture is:

Lexical Retrieval
+
Dense Retrieval
↓
Fusion
↓
Hybrid Candidate Set
↓
Reranking
↓
Final Results

Metadata Filtering

Hybrid retrieval can also operate together with structured constraints.

Lexical Evidence
+
Semantic Evidence
+
Metadata Filters
↓
More Targeted Retrieval

Query Routing

Query routing chooses or adjusts retrieval strategies based on the query.

Query
↓
Analyze Query Type
↓
Choose Retrieval Strategy

This is related to hybrid retrieval but is not the same concept.

Hybrid Retrieval Does Not Guarantee Truth

Lexical Match
+
Semantic Match
≠
Factual Accuracy

Hybrid retrieval improves relevance and candidate coverage, not factual verification.

The complete Chapter 10 pipeline is:

Corpus
↓
Lexical Index + Vector Index
↓
User Query
├───────────────┬────────────────┐
↓               ↓
Lexical         Dense
Retrieval       Retrieval
↓               ↓
Candidates      Candidates
└───────┬───────┘
        ↓
      Fusion
        ↓
Hybrid Candidate Set
        ↓
Ranking / Reranking
        ↓
Final Results

We now understand how multiple retrievers can work together.

But candidate generation is only part of the retrieval system.

The next challenge is deciding:

Which candidate is actually best?

and:

How should the final result set be ordered?

That leads directly into Chapter 11:

Ranking and Reranking.

The Single Most Important Idea from Chapter 10

Hybrid retrieval combines lexical evidence and semantic evidence so that a system can capture both exact term matches and meaning-based matches. The retrieved candidates are then fused—using score-based or rank-based methods such as weighted fusion or RRF—and can be passed to later ranking or reranking stages for more precise ordering.