Core lesson

Query Expansion

Improve retrieval by rewriting, expanding, or generating related query forms so the search system can find evidence users did not phrase exactly.

3 min readUpdated July 2026

After this, you will understand

How Query Expansion helps you see how prompts, tools, retrieval, agents, and evals become real AI engineering workflows.

Article guideprerequisites, mental models, and concepts

intermediateRetrievalMechanicsEvaluation

Three useful mental models

In plain terms

Treat the idea as a definition to memorize.

Production pressure

Real systems force the idea to handle Query Expansion, Query Rewriting, and HyDE.

Better reasoning

Use the concept to decide what the system guarantees, what it risks, and what it costs to operate.

Think before reading

Where would Query Expansion appear in a real production system, and what failure or bottleneck would it help you reason about?

As you read, look for the pressure that creates the idea first. The mechanics matter more once the reason is clear.

Connected learning

These lessons add useful context to the current core lesson.

Concepts Covered

Query expansion
Query rewriting
Multi-query retrieval
HyDE
Recall improvement
Retrieval drift
Search intent
RAG query preprocessing

Definition

Query expansion is the practice of transforming a user's query into one or more richer search queries before retrieval.

The plain-English version:

help the retrieval system search for what the user means, not only what they typed

Expansion can add synonyms, rewrite vague wording, generate multiple query variants, or create a hypothetical answer-shaped document for embedding.

Why This Concept Exists

Users rarely phrase questions the same way documents phrase answers.

A user may ask:

Can I get money back after renewal?

The policy may say:

Annual plan refunds are available within 30 days of renewal.

The retrieval system has to connect "money back" with "refund" and "after renewal" with "renewal policy."

Query expansion exists because short, vague, or user-shaped questions can be weak retrieval inputs.

The Beginner Mental Model

A beginner may think:

The best query is always the exact user query.

Sometimes it is. But in many RAG systems, the user query is not shaped like the corpus.

A better mental model:

user query -> retrieval-oriented query forms -> candidate search

The goal is not to change the user's question. The goal is to improve recall while preserving intent.

Common Expansion Strategies

Query expansion can happen in several ways.

Synonym expansion: add related terms such as "refund" for "money back."

Query rewriting: turn conversation context into a standalone search query.

Multi-query retrieval: generate several query variants and merge results.

HyDE: generate a hypothetical document or answer-shaped passage, embed that, and retrieve real documents near it.

Step-back style rewriting: search for a more general concept before narrowing down.

Each strategy tries to reduce mismatch between user wording and corpus wording.

A Concrete Example

In a chat session, the user asks:

What about contractors?

That query is useless without context.

The previous turn was about:

home office equipment reimbursement

A query rewrite could produce:

contractor home office equipment reimbursement policy

That expanded query is much better for retrieval.

The model did not answer yet. It only made the retrieval request more useful.

Retrieval Drift

Expansion can improve recall, but it can also drift.

If the expansion adds concepts the user did not ask about, retrieval may move toward the wrong evidence.

For example:

user query: "refund after renewal"
bad expansion: "chargeback dispute and cancellation fraud"

The expanded query may retrieve related but wrong policy areas.

This is why query expansion should be evaluated against real questions and failure cases.

Where Query Expansion Fits

Query expansion usually happens before candidate retrieval:

user question
  -> rewrite or expand
  -> retrieve candidates
  -> merge and deduplicate
  -> rerank
  -> assemble context

For multi-query systems, each query variant may retrieve its own candidates before merging.

For HyDE-style systems, the generated hypothetical text becomes the embedding input for retrieval.

Common Confusions

Query expansion is not answering the question.

It prepares retrieval. The final answer still needs grounded context.

Expansion is not always better.

Bad expansion can retrieve confident nonsense.

HyDE does not require the hypothetical document to be factually correct.

Its purpose is to create a retrieval representation that points toward likely relevant real documents.

Conversation rewriting is not memory.

It uses available context to make the current query searchable.

What This Does Not Mean

Query expansion does not remove the need for filters, reranking, or answer grounding.

It improves the query side of retrieval. The system still needs to verify that retrieved documents actually support the answer.

Finished reading?

Your reading history is saved in this browser so you can continue later.

Recommended Next

Parent-Child RetrievalBuilding With AI3 min read

This turns the foundation vocabulary into a practical AI engineering decision.

Optional exploration

These links add context, but they do not replace the recommended next lesson.