Chapter 16- Query Understanding

Chapter Goal

In Chapter 15, we prepared information for retrieval:

Documents
↓
Parsing
↓
Chunking
↓
Embedding
↓
Indexing
↓
Searchable Collection

Now the collection is ready.

But another problem remains.

The user's original query is not always the best possible query for retrieval.

A user might ask:

How can I stop paying for my plan every year?

while the indexed documents may use terms such as:

cancel annual subscription
disable automatic renewal
terminate yearly membership

Or the user may ask:

What are the requirements for Dutch citizenship
and how long do I need to live there?

This is actually asking more than one question.

Modern retrieval systems can therefore process the user query before searching.

The broad process is called:

Query Understanding

The system may:

interpret the query
identify its intent
rewrite it
expand it
split it into subqueries
generate multiple related searches
retrieve results from several queries

The purpose is not to replace the user's information need.

The purpose is to represent that need in forms that make retrieval more effective.

The central idea is:

User Query
↓
Query Understanding
↓
Better Retrieval Queries
↓
Retriever
↓
Relevant Information

By the end of this chapter, you should understand:

• what Query Understanding means;
• what Search Intent is;
• what Intent Classification does;
• what Query Rewriting is;
• what Query Reformulation means;
• what Query Expansion is;
• how Synonym Expansion works;
• what Semantic Expansion adds;
• what Query Decomposition is;
• what Subqueries are;
• what Query Fan-Out means;
• why systems may create Multiple Search Queries;
• what Parallel Retrieval means;
• what Multi-Query Retrieval is;
• what Query Variants are;
• and how these techniques differ from one another.

The chapter's full conceptual pipeline is:

User Query
↓
Query Interpretation
↓
Rewriting / Expansion / Decomposition
↓
Query Variants or Subqueries
↓
Retrieval
↓
Combined Relevant Results

16.1 Query Interpretation

1. What Is Query Understanding?

Query Understanding is the process of interpreting what the user's query means and what information the user is trying to obtain.

Consider:

best Persian dentist Toronto

The system should not treat this only as four unrelated words.

The query expresses something closer to:

Find
↓
Dentists

Preference
↓
Persian / Persian-speaking

Location
↓
Toronto

Decision Goal
↓
Compare good options

Query Understanding tries to recover this underlying information need.

Query Text
≠
Complete Information Need

The query is the user's expression of that need.

2. Why Query Understanding Matters

A retrieval system can be technically excellent and still return poor results if it searches for the wrong interpretation.

Suppose the query is:

Apple repair

Possible meanings include:

Apple device repair
apple tree repair
fruit processing equipment repair

If the intended meaning is:

Apple Inc. device repair

but the system interprets:

apple
→ fruit

then even a strong retriever may search the wrong information space.

Wrong Interpretation
↓
Wrong Retrieval Target
↓
Relevant Results for the Wrong Meaning

So query understanding happens upstream of retrieval quality.

3. Query Interpretation

Query Interpretation means determining the most likely meaning of the query in its current context.

For example:

python installation

likely means:

Python programming language

while:

python habitat

likely means:

python snake

The surrounding terms help determine the intended meaning.

Query Terms
+
Context
↓
Likely Meaning

4. Query Interpretation Can Use Context

Imagine this conversation:

User:
Tell me about hybrid retrieval.

User:
How does it improve recall?

The second query alone is incomplete:

How does it improve recall?

But conversation context tells us:

it
→ hybrid retrieval

The effective information need is:

How does hybrid retrieval improve recall?

So query interpretation can depend on:

current query
+
conversation context

5. What Is Search Intent?

Search Intent describes what the user is trying to accomplish with the query.

Compare:

what is RAG

with:

RAG implementation tutorial

and:

best RAG platform

All three relate to RAG.

But the user's goals differ.

what is RAG
→ explanation

RAG implementation tutorial
→ implementation guidance

best RAG platform
→ comparison / selection

A good retrieval system should recognize that different intents require different information.

6. Common Search Intent Patterns

Useful intent patterns can include:

informational
navigational
transactional
local
comparative

For example:

What is BM25?
→ informational

OpenAI API documentation
→ navigational

buy running shoes
→ transactional

dentist near me
→ local

Semrush vs. Ahrefs
→ comparative

These labels are useful abstractions.

Real queries can contain more than one intent.

7. Intent Can Be Mixed

Consider:

best immigration lawyers Toronto and their prices

This query contains several goals:

find lawyers
+
compare options
+
find pricing

It is therefore not always useful to force every query into exactly one intent category.

A query may express:

Primary Intent
+
Secondary Intent

or several related needs.

8. What Is Intent Classification?

Intent Classification is the process of assigning a query to one or more intent categories.

User Query
↓
Intent Classifier
↓
Intent Category

For example:

dentist near me
↓
Local Intent

or:

how does dense retrieval work
↓
Informational Intent

Intent classification can help the system decide:

which sources to search
which retrieval strategy to emphasize
which ranking signals matter

9. Intent Classification Is Not Perfect Understanding

Intent classification simplifies the query.

For example:

best Persian dentist Toronto

might be labeled:

Local + Comparative

But this label does not capture everything.

The query also contains:

Service:
dentist

Preference:
Persian

Location:
Toronto

Intent Classification
⊂
Query Understanding

Intent classification is one part of query understanding, not the whole process.

10. Interpretation Before Retrieval

At the simplest level:

Raw User Query
↓
Interpret Meaning
↓
Identify Intent
↓
Create Better Search Representation
↓
Retrieval

This gives the retriever a clearer representation of what the user actually needs.

16.2 Query Rewriting

11. What Is Query Rewriting?

Query Rewriting means transforming the original user query into a different query that is more useful for retrieval while preserving the original information need.

Example:

Original Query:
How do I stop paying every year?

Possible rewrite:

cancel annual subscription automatic renewal

The language changed.

The user's goal did not.

Original Information Need
=
Rewritten Information Need

while:

Original Wording
≠
Rewritten Wording

12. Why Rewrite a Query?

Users do not necessarily use the same vocabulary as the documents.

Suppose the user asks:

How do I stop my plan from renewing?

but the documentation says:

Disable Automatic Renewal

A rewrite can produce:

disable automatic renewal subscription

This reduces the vocabulary gap between:

User Language
and
Document Language

13. Query Rewriting Example

Original query:

Can I get my money back if I cancel?

Possible retrieval rewrite:

subscription cancellation refund policy

The rewrite makes important concepts explicit:

subscription
cancellation
refund
policy

14. Rewriting Can Make Conversational Queries Self-Contained

Suppose the conversation is:

User:
Tell me about BM25.

User:
How is it different from dense retrieval?

The second query depends on conversation history.

A rewrite can create:

How is BM25 different from dense retrieval?

Context-Dependent Query
+
Conversation History
↓
Self-Contained Query

15. What Is Query Reformulation?

Query Reformulation is the broader idea of expressing the same information need in a different query form.

For example:

Original:
How can I stop paying every year?

Reformulated:

How can I cancel annual subscription renewal?

Another reformulation:

disable recurring yearly subscription payment

All represent approximately the same underlying need.

16. Good Rewriting Preserves Intent

A good rewrite may make the query clearer.

It should not silently change important constraints.

Original:

Persian dentist Toronto

Bad rewrite:

dentist Canada

The rewrite removed:

Persian
Toronto

which were important parts of the information need.

Rewrite
→ improve expression

not

Rewrite
→ change the user's goal

17. Rewriting Can Also Fail

A rewritten query can introduce information the user never provided.

Original:

Persian dentist

Incorrect rewrite:

Persian dentist Toronto

If Toronto was never stated or implied, the system has invented a location constraint.

Original Query
↓
Bad Rewrite
↓
Wrong Retrieval

Rewriting should remain grounded in the user's actual query and available context.

16.3 Query Expansion

18. What Is Query Expansion?

Query Expansion adds related terms or concepts to the original query to increase retrieval coverage.

Suppose the query is:

car repair

An expanded version could include:

car repair
automobile repair
auto mechanic
vehicle repair

Original Query
+
Related Terms
↓
Expanded Query

19. Query Rewriting vs. Query Expansion

Query Rewriting

How do I stop paying every year?
↓
cancel annual subscription

Query Expansion

cancel annual subscription
+
terminate yearly membership
+
disable recurring billing

Rewriting
→ reformulates

Expansion
→ broadens useful retrieval signals

20. Synonym Expansion

Synonym Expansion adds words or phrases with similar meaning.

Example:

lawyer
+
attorney

Another example:

cheap
+
affordable
+
low-cost

The expanded query gives lexical retrieval more opportunities to match relevant documents.

21. Why Synonym Expansion Helps Lexical Retrieval

Suppose the user asks:

cheap immigration lawyer

but a relevant page says:

affordable immigration attorney

A pure exact-term system may have weak overlap.

Expansion creates:

cheap
affordable

lawyer
attorney

Now lexical matching has stronger signals.

22. Synonyms Are Context-Dependent

Synonyms cannot always be added blindly.

Consider:

Java developer

The word:

Java

could also refer to:

island
coffee

but those meanings are irrelevant here.

Expansion
must respect
Query Context

23. What Is Semantic Expansion?

Semantic Expansion goes beyond direct synonyms.

It adds concepts that are meaningfully related to the information need.

For example:

cancel subscription

might semantically expand toward:

automatic renewal
recurring billing
membership termination
subscription management

These are not all strict synonyms.

But they belong to the same semantic problem.

Synonym Expansion
→ similar words

Semantic Expansion
→ related concepts

24. Semantic Expansion Example

Query:

MacBook battery problem

Semantic expansion might include:

battery health
battery replacement
charging issue
battery service
MacBook power problem

The system is exploring the conceptual neighborhood of the query.

This can improve coverage when relevant documents describe the problem differently.

25. Expansion Can Improve Recall

Suppose the original query matches only:

Document A
Document B

After useful expansion:

Document A
Document B
Document C
Document D

may become retrievable.

More Useful Query Expressions
↓
More Relevant Candidates Found
↓
Potentially Higher Recall

26. Expansion Can Introduce Noise

If expansion becomes too broad:

Original Intent
↓
Too Many Related Concepts
↓
Irrelevant Candidates

For example:

Apple repair

should not expand into:

fruit
orchard
tree
agriculture

if the context clearly refers to Apple devices.

More Expansion
≠
Automatically Better Retrieval

16.4 Query Decomposition

27. What Is Query Decomposition?

Some user questions contain several distinct information needs.

Query Decomposition breaks a complex query into smaller queries.

Example:

What are the requirements for Dutch citizenship,
how long do I need to live in the Netherlands,
and what language exam is required?

This can become:

Subquery 1:
Dutch citizenship eligibility requirements

Subquery 2:
Required residence period for Dutch citizenship

Subquery 3:
Language exam requirement for Dutch citizenship

Each subquery can then be retrieved separately.

28. Why Decompose a Query?

A single document may not answer every part of a complex question.

Question
├── eligibility
├── residence duration
└── language requirement

Relevant evidence might exist in:

Document A
→ eligibility

Document B
→ residence duration

Document C
→ language requirement

Decomposition allows each information need to search independently.

29. What Is a Subquery?

A Subquery is one smaller retrieval query created from a larger query.

Original:

Compare ChatGPT and Gemini for research
and tell me which one supports citations better.

Possible subqueries:

Subquery 1:
ChatGPT research capabilities

Subquery 2:
Gemini research capabilities

Subquery 3:
ChatGPT citation support

Subquery 4:
Gemini citation support

The results can later be combined.

30. Good Subqueries Should Preserve the Original Task

Decomposition should not create unrelated questions.

The purpose is:

Complex Information Need
↓
Smaller Connected Information Needs

All subqueries should contribute to answering the original question.

Subqueries
→ pieces of original task

not

Subqueries
→ random related searches

31. Decomposition Is Especially Useful for Multi-Part Questions

Simple query:

What is BM25?

does not need meaningful decomposition.

Complex query:

How does BM25 differ from dense retrieval,
when should each be used,
and how can they be combined?

can naturally become:

Q1:
BM25 vs. dense retrieval

Q2:
When to use BM25

Q3:
When to use dense retrieval

Q4:
How BM25 and dense retrieval can be combined

Decomposition is therefore task-dependent.

16.5 Query Fan-Out

32. What Is Query Fan-Out?

Query Fan-Out means generating multiple search queries from one original user query and sending them outward into retrieval.

Original Query
↓
Query Fan-Out
├── Search Query A
├── Search Query B
├── Search Query C
└── Search Query D

Each query can retrieve different evidence.

The results can then be brought back together.

33. Why Fan Out a Query?

A single query represents only one wording or angle.

Suppose the user asks:

Why is my website not appearing in AI answers?

The system might fan this out into:

AI search source discovery problems

website retrieval eligibility AI search

brand citation visibility AI answers

passage relevance generative search

Each query explores a different angle of the information need.

The goal is broader evidence discovery.

34. Query Fan-Out vs. Query Decomposition

Query Decomposition

Complex Question
↓
Subproblem A
Subproblem B
Subproblem C

Query Fan-Out

Original Need
↓
Query A
Query B
Query C

Decomposition
→ divide the task

Fan-Out
→ broaden the search paths

They can also be used together.

35. Multiple Search Queries

Query Fan-Out produces:

Multiple Search Queries

For example:

Original:
best Persian restaurants Toronto

Possible queries:

Persian restaurants Toronto

Iranian restaurants Toronto

best Iranian food Toronto

Persian dining Toronto

These queries may retrieve overlapping results, but they may also expose sources that one formulation misses.

36. Why Multiple Queries Can Help

Different documents use different terminology.

One source may use:

Persian restaurant

another:

Iranian restaurant

another:

Iranian cuisine

Multiple queries increase the chance of covering those variations.

Query A
→ Candidate Set A

Query B
→ Candidate Set B

Query C
→ Candidate Set C

Then:

A + B + C
↓
Broader Candidate Pool

37. Parallel Retrieval

Parallel Retrieval means multiple search queries can be executed independently rather than one after another.

                 ┌→ Query A → Retrieval A
Original Query ──┼→ Query B → Retrieval B
                 └→ Query C → Retrieval C

Then:

Results A
+
Results B
+
Results C
↓
Combined Results

The queries do not necessarily need to wait for one another.

38. Parallel Does Not Mean the Queries Are Identical

Each branch can search for something slightly different.

For example:

Query A:
Persian dentist Toronto

Query B:
Iranian dental clinic Toronto

Query C:
Farsi-speaking dentist Toronto

Each branch may discover different candidates.

The value comes from diversity in the search expressions.

39. Fan-Out Can Increase Both Coverage and Noise

More search queries can produce more relevant results.

But they can also produce more irrelevant candidates.

More Queries
↓
Potentially More Coverage
+
Potentially More Noise

Therefore downstream stages such as:

ranking
reranking
deduplication

may become important after fan-out.

The goal is not simply:

generate as many searches as possible

but:

generate useful complementary searches

16.6 Multi-Query Retrieval

40. What Is Multi-Query Retrieval?

Multi-Query Retrieval is a retrieval approach in which several query formulations are used to retrieve information for the same overall information need.

Original Query
↓
Multiple Query Variants
↓
Multiple Retrieval Runs
↓
Candidate Results
↓
Combined Retrieval Set

Instead of relying on one query representation:

Q1

the system may retrieve with:

Q1
Q2
Q3
Q4

and combine the results.

41. What Is a Query Variant?

A Query Variant is an alternative formulation of the same or closely related information need.

Original query:

How do I cancel automatic renewal?

Variants:

disable subscription auto-renewal

stop recurring subscription payment

turn off automatic membership renewal

The wording differs.

The underlying need remains similar.

42. Why Query Variants Matter

Suppose a relevant document is optimized around:

disable auto-renewal

while the user asks:

stop paying every year

One query may miss the document.

Another variant may match it strongly.

One Information Need
↓
Several Query Representations
↓
Different Retrieval Opportunities

43. Multi-Query Retrieval Example

Original:

How can I cancel my annual subscription?

The system generates:

Q1:
cancel annual subscription

Q2:
disable automatic renewal

Q3:
terminate yearly membership

Q4:
stop recurring billing

Each query retrieves candidates:

Q1 → A, B, C

Q2 → B, D, E

Q3 → A, F

Q4 → D, G

The combined pool becomes:

A
B
C
D
E
F
G

Duplicates can be removed.

Ranking or reranking can then identify the strongest candidates.

44. Multi-Query Retrieval vs. Query Expansion

Query Expansion

Creates one richer query representation:

car repair
+
auto mechanic
+
vehicle repair

Multi-Query Retrieval

Creates separate queries:

Q1: car repair

Q2: auto mechanic

Q3: vehicle repair

and executes retrieval for each.

Query Expansion
→ expand one query

Multi-Query Retrieval
→ retrieve using multiple queries

45. Multi-Query Retrieval vs. Fan-Out

Query Fan-Out describes the process of branching one user query into multiple search queries.

Multi-Query Retrieval describes the retrieval strategy that uses those multiple queries to collect candidates.

Query Fan-Out
→ create branches

Multi-Query Retrieval
→ search using those branches

For example:

Original Query
↓
Fan-Out
├── Q1
├── Q2
└── Q3
↓
Multi-Query Retrieval
├── Results 1
├── Results 2
└── Results 3
↓
Combined Candidate Set

46. Query Decomposition vs. Fan-Out vs. Multi-Query Retrieval

Query Decomposition

Complex Task
↓
Different Subproblems

Example:

Eligibility?
Residence period?
Language requirement?

Query Fan-Out

One Need
↓
Multiple Search Paths

Example:

Persian dentist Toronto
Iranian dentist Toronto
Farsi-speaking dentist Toronto

Multi-Query Retrieval

Multiple Queries
↓
Multiple Retrieval Runs
↓
Combined Candidates

So:

Decomposition
→ what smaller questions should we ask?

Fan-Out
→ what search queries should we generate?

Multi-Query Retrieval
→ how do we retrieve with all those queries?

47. Query Processing Does Not Replace Retrieval

Everything in this chapter happens around the query.

But these techniques do not themselves search the collection.

For example:

Query Rewriting
→ creates a better query

Then:

Retriever
→ searches the index with that query

Similarly:

Query Fan-Out
→ creates several queries

Then:

Retrieval
→ runs searches for them

Query Processing
≠
Retrieval

Instead:

Query Processing
→ prepares retrieval requests

48. A Complete Example

Suppose the user asks:

I'm moving to Toronto.
Can you find Persian-speaking dentists
and tell me which ones are open on Saturday?

Step 1 — Query Interpretation

The system identifies:

Service:
dentist

Language Preference:
Persian-speaking

Location:
Toronto

Availability:
Saturday

Goal:
find suitable providers

Step 2 — Query Rewriting

A cleaner retrieval query might be:

Persian-speaking dentists in Toronto open Saturday

Step 3 — Query Expansion

Useful related concepts might include:

Iranian dentist Toronto
Farsi-speaking dentist Toronto
Persian dental clinic Toronto

Step 4 — Query Fan-Out

The system creates:

Q1:
Persian-speaking dentist Toronto Saturday

Q2:
Iranian dentist Toronto open Saturday

Q3:
Farsi-speaking dental clinic Toronto weekend hours

Step 5 — Parallel Retrieval

Q1 → Retrieval
Q2 → Retrieval
Q3 → Retrieval

Each produces candidates.

Step 6 — Multi-Query Retrieval

The candidates are combined:

Results Q1
+
Results Q2
+
Results Q3
↓
Combined Candidate Pool

The system can then rank or rerank the results.

The full query-side pipeline becomes:

User Query
↓
Interpretation
↓
Rewrite
↓
Expansion
↓
Fan-Out
↓
Multiple Queries
↓
Parallel Retrieval
↓
Combined Candidates

49. Common Misunderstandings

Misunderstanding 1: Query Understanding is the same as Intent Classification

Incorrect. Intent Classification is only one component of Query Understanding.

Misunderstanding 2: The user's exact query must always be sent directly to retrieval

Incorrect. The system may rewrite, expand, decompose, or generate variants first.

Misunderstanding 3: Query Rewriting means changing the user's goal

Incorrect. A good rewrite preserves the information need while improving its formulation.

Misunderstanding 4: Query Reformulation always produces only one alternative query

Not necessarily. The same need can have several reformulations.

Misunderstanding 5: Query Expansion only means adding synonyms

Incorrect. Semantic Expansion can also add related concepts.

Misunderstanding 6: More expansion always improves retrieval

Incorrect. Poor expansion can introduce noise or move away from the original intent.

Misunderstanding 7: Query Decomposition means creating synonyms

Incorrect. Decomposition splits a complex information need into smaller subproblems.

Misunderstanding 8: Every query should be decomposed

Incorrect. Simple queries often do not need decomposition.

Misunderstanding 9: Query Fan-Out and Query Decomposition are identical

Incorrect. Decomposition divides a task into subproblems; fan-out creates multiple retrieval paths.

Misunderstanding 10: Parallel Retrieval means all queries have the same wording

Incorrect. The queries can be different and complementary.

Misunderstanding 11: Query Expansion and Multi-Query Retrieval are the same

Incorrect. Expansion enriches a query; Multi-Query Retrieval performs retrieval using several queries.

Misunderstanding 12: Query Fan-Out and Multi-Query Retrieval are exactly the same stage

Not quite. Fan-Out creates multiple queries. Multi-Query Retrieval uses them to retrieve candidates.

Misunderstanding 13: Query processing guarantees good results

Incorrect. The queries can still be poorly interpreted or transformed, and retrieval can still fail.

50. How the Concepts Connect

Chapter 15 prepared the information side:

Documents
↓
Parsing
↓
Chunking
↓
Embedding
↓
Indexing
↓
Searchable Collection

Chapter 16 prepares the query side:

User Query
↓
Query Understanding
↓
Interpretation
↓
Rewriting / Expansion / Decomposition
↓
Query Variants
↓
Retrieval

Now both sides can meet:

QUERY SIDE

User Query
↓
Query Understanding
↓
Search Queries

        ↓

RETRIEVAL

        ↓

INDEX SIDE

Searchable Documents / Chunks

The complete architecture is becoming:

SOURCE PREPARATION

Documents
↓
Parsing
↓
Chunking
↓
Embedding
↓
Indexing
↓
Searchable Collection


QUERY PREPARATION

User Query
↓
Interpretation
↓
Rewriting
↓
Expansion
↓
Decomposition / Fan-Out
↓
Query Variants


RETRIEVAL

Search Queries
+
Searchable Collection
↓
Retriever
↓
Relevant Candidates

We now understand:

Chapter 13
→ Why retrieval is needed

Chapter 14
→ How retrieval connects to generation

Chapter 15
→ How source information is prepared

Chapter 16
→ How the user query is prepared

The next question is:

Once retrieval returns several passages, which ones should actually enter the LLM's limited context window?

That leads into:

Chapter 17 — Building the Context for the LLM

Mental Model

Imagine someone walks into a large library and says:

I need information about how to stop
my membership from charging me every year.

The librarian does not necessarily search those exact words.

First, the librarian interprets the request:

The user wants to cancel
automatic annual renewal.

That is:

Query Interpretation

The librarian converts the request into more useful catalog language:

annual subscription cancellation
automatic renewal

That is:

Query Rewriting

Then the librarian considers related terminology:

recurring billing
membership termination
disable auto-renewal

That is:

Query Expansion

If the question contains several parts, the librarian separates them:

How do I cancel?
When does cancellation take effect?
Can I receive a refund?

That is:

Query Decomposition

The librarian may then search several formulations:

cancel annual subscription

disable auto-renewal

stop recurring membership billing

That is:

Query Fan-Out

Several catalog searches run:

Query A → Results A
Query B → Results B
Query C → Results C

That is:

Multi-Query Retrieval

The simplest mental model is:

Interpret
→ understand what the user wants

Rewrite
→ express it better for search

Expand
→ add useful related concepts

Decompose
→ split complex needs

Fan Out
→ create multiple search paths

Multi-Query Retrieval
→ search those paths and combine results

And the simplest reliable definition is:

Query Understanding is the process of interpreting a user's information need and transforming the original query into retrieval-ready forms. Depending on the task, the system may rewrite the query, expand its vocabulary or concepts, decompose complex questions into subqueries, fan one query out into several search paths, and retrieve across multiple query variants.

Chapter 16 Recap

Query Understanding

The broad process of interpreting and preparing the user's query for retrieval.

User Query
↓
Understanding
↓
Better Retrieval Representation

Query Interpretation

Determining the likely meaning of the query.

Query
+
Context
↓
Likely Meaning

Search Intent

What the user is trying to accomplish.

informational
navigational
transactional
local
comparative

Intent Classification

Assigning the query to one or more intent categories.

Query
↓
Intent Classification
↓
Intent Category

It is one part of Query Understanding.

Query Rewriting

Transforming the query into a more retrieval-friendly form while preserving the information need.

Original Query
↓
Rewrite
↓
Better Search Query

Query Reformulation

Expressing the same information need differently.

Information Need
↓
Alternative Query Form

Query Expansion

Adding useful related terminology or concepts.

Original Query
+
Related Terms
↓
Expanded Query

Synonym Expansion

Adding words or phrases with similar meaning.

lawyer
+
attorney

Semantic Expansion

Adding conceptually related ideas rather than only strict synonyms.

cancel subscription
↓
automatic renewal
recurring billing
membership termination

Query Decomposition

Breaking a complex query into smaller retrieval problems.

Complex Query
↓
Subquery 1
Subquery 2
Subquery 3

Subquery

One smaller search question created from a larger information need.

Query Fan-Out

Creating multiple search queries from one original query.

Original Query
↓
Q1
Q2
Q3

Multiple Search Queries

Different formulations or search paths used to explore the information need.

Parallel Retrieval

Running multiple retrieval operations independently.

Q1 → Retrieval 1
Q2 → Retrieval 2
Q3 → Retrieval 3

Multi-Query Retrieval

Using several queries to retrieve candidates for the same overall information need.

Query Variants
↓
Multiple Retrieval Runs
↓
Combined Candidates

Query Variant

An alternative formulation of the same or closely related need.

cancel annual subscription

disable automatic renewal

stop recurring billing

Query Expansion vs. Multi-Query Retrieval

Query Expansion
→ enrich one query

Multi-Query Retrieval
→ retrieve using multiple queries

Query Decomposition vs. Query Fan-Out

Decomposition
→ split the problem

Fan-Out
→ create multiple search paths

Query Fan-Out vs. Multi-Query Retrieval

Fan-Out
→ create multiple queries

Multi-Query Retrieval
→ execute retrieval with them

The complete Chapter 16 pipeline is:

User Query
↓
Query Understanding
↓
Query Interpretation
↓
Search Intent / Intent Classification
↓
Query Rewriting / Reformulation
↓
Query Expansion
├── Synonym Expansion
└── Semantic Expansion
↓
Query Decomposition
↓
Subqueries
↓
Query Fan-Out
↓
Multiple Search Queries
↓
Parallel Retrieval
↓
Multi-Query Retrieval
↓
Combined Candidate Results

We now have both sides of the retrieval system:

DOCUMENT SIDE

Document
→ Parsing
→ Chunking
→ Embedding
→ Indexing


QUERY SIDE

Query
→ Understanding
→ Rewriting
→ Expansion
→ Fan-Out

        ↓

Retrieval

The next step is:

Chapter 17 — Building the Context for the LLM

The Single Most Important Idea from Chapter 16

The user's original wording is not always the best representation of the user's information need for retrieval. Query Understanding interprets that need and can transform it through rewriting, expansion, decomposition, and multiple query variants so the retrieval system has more effective ways to find the relevant information without changing what the user actually wants to know.