Definition
Query expansion refers to techniques that enhance search queries by adding additional terms, synonyms, or alternative phrasings to improve retrieval recall. In traditional search, this meant adding synonyms from WordNet or co-occurring terms from the corpus. In modern RAG systems, LLMs can generate multiple query variants, hypothetical answers, or decompose complex questions. The goal is bridging the vocabulary gap between how users phrase questions and how relevant documents are written, ultimately retrieving more relevant context.
Why it matters
Query expansion addresses fundamental search challenges:
- Vocabulary mismatch — users say “car” but documents say “vehicle” or “automobile”
- Query ambiguity — short queries lack context, expansion clarifies intent
- Improved recall — find relevant documents that exact queries miss
- RAG enhancement — better retrieval leads to better LLM responses
- Multi-faceted queries — decompose complex questions into searchable parts
- Underspecified queries — add implicit context users didn’t express
Without query expansion, retrieval systems miss relevant content due to surface-level lexical mismatches.
How it works
┌────────────────────────────────────────────────────────────┐
│ QUERY EXPANSION │
├────────────────────────────────────────────────────────────┤
│ │
│ THE PROBLEM: │
│ ──────────── │
│ │
│ User query: "cheap flights to NYC" │
│ │
│ Documents might say: │
│ • "affordable airfare to New York" │
│ • "budget airline tickets to JFK" │
│ • "low-cost travel to Manhattan" │
│ │
│ Exact match misses all of these! │
│ │
│ │
│ CLASSICAL QUERY EXPANSION METHODS: │
│ ────────────────────────────────── │
│ │
│ 1. SYNONYM EXPANSION (WordNet/Thesaurus) │
│ ───────────────────────────────────────── │
│ │
│ Original: "cheap flights NYC" │
│ ↓ ↓ ↓ │
│ Expanded: │
│ • cheap → inexpensive, affordable, budget, low-cost │
│ • flights → airfare, air travel, plane tickets │
│ • NYC → New York, New York City, JFK, LaGuardia │
│ │
│ 2. CORPUS-BASED EXPANSION │
│ ───────────────────────── │
│ │
│ Find terms that co-occur with query terms: │
│ │
│ "flights" often appears with: │
│ [airline, booking, departure, arrival, layover] │
│ │
│ Add statistically associated terms │
│ │
│ 3. PSEUDO-RELEVANCE FEEDBACK (PRF) │
│ ────────────────────────────────── │
│ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ 1. Run initial query │ │
│ │ 2. Assume top-k results are relevant │ │
│ │ 3. Extract terms from those documents │ │
│ │ 4. Add extracted terms to query │ │
│ │ 5. Re-run expanded query │ │
│ └─────────────────────────────────────────────────┘ │
│ │
│ Original query: "machine learning" │
│ Top results contain: neural, network, training, model│
│ Expanded: "machine learning neural network training" │
│ │
│ │
│ MODERN LLM-BASED EXPANSION: │
│ ─────────────────────────── │
│ │
│ 1. MULTI-QUERY GENERATION │
│ ───────────────────────── │
│ │
│ Prompt: "Generate 3 alternative versions of this │
│ question that might retrieve relevant │
│ documents: {original_query}" │
│ │
│ Original: "What causes diabetes?" │
│ │
│ Generated queries: │
│ • "Risk factors and etiology of diabetes mellitus" │
│ • "How does insulin resistance develop?" │
│ • "Genetic and lifestyle causes of type 2 diabetes" │
│ │
│ ┌─────────────────────────────────────────────────┐ │
│ │ Search each variant → Union of results │ │
│ │ OR │ │
│ │ Embed all variants → Use mean embedding │ │
│ └─────────────────────────────────────────────────┘ │
│ │
│ │
│ 2. HyDE: HYPOTHETICAL DOCUMENT EMBEDDINGS │
│ ───────────────────────────────────────── │
│ │
│ Instead of expanding query, generate hypothetical │
│ answer document and search for similar real docs: │
│ │
│ Query: "How does CRISPR work?" │
│ ↓ │
│ LLM generates hypothetical answer: │
│ ┌──────────────────────────────────────────────────┐ │
│ │ "CRISPR (Clustered Regularly Interspaced Short │ │
│ │ Palindromic Repeats) is a gene editing │ │
│ │ technology that uses Cas9 protein to cut DNA │ │
│ │ at specific locations. A guide RNA directs │ │
│ │ Cas9 to the target sequence where it makes │ │
│ │ a double-strand break..." │ │
│ └──────────────────────────────────────────────────┘ │
│ ↓ │
│ Embed this hypothetical document │
│ ↓ │
│ Search for real documents similar to it │
│ │
│ Why this works: Hypothetical answer is in "document │
│ space" not "query space" - closer to actual docs │
│ │
│ │
│ 3. STEP-BACK PROMPTING │
│ ────────────────────── │
│ │
│ Generate more abstract/general queries: │
│ │
│ Original: "Why did Apple stock drop on Jan 15 2024?" │
│ ↓ │
│ Step-back: "What factors affect Apple stock price?" │
│ "Apple earnings and revenue trends" │
│ │
│ Retrieves context that helps answer specific Q │
│ │
│ │
│ 4. QUERY DECOMPOSITION │
│ ────────────────────── │
│ │
│ Break complex query into sub-queries: │
│ │
│ Original: "How does Tesla's FSD compare to Waymo's │
│ approach, and which is safer?" │
│ ↓ │
│ Sub-queries: │
│ • "How does Tesla Full Self-Driving work?" │
│ • "How does Waymo autonomous driving work?" │
│ • "Tesla FSD safety statistics and accidents" │
│ • "Waymo safety record and statistics" │
│ │
│ Retrieve for each, then synthesize answer │
│ │
│ │
│ RAG PIPELINE WITH QUERY EXPANSION: │
│ ────────────────────────────────── │
│ │
│ User Query │
│ │ │
│ ↓ │
│ ┌────────────┐ │
│ │ Expand │──→ Generate variants / HyDE / Decompose│
│ └────────────┘ │
│ │ │
│ ↓ │
│ ┌────────────┐ │
│ │ Retrieve │──→ Search with all expanded queries │
│ └────────────┘ │
│ │ │
│ ↓ │
│ ┌────────────┐ │
│ │ Dedupe & │──→ Remove duplicate chunks │
│ │ Rerank │──→ Score by relevance to original │
│ └────────────┘ │
│ │ │
│ ↓ │
│ ┌────────────┐ │
│ │ Generate │──→ LLM answers with expanded context │
│ └────────────┘ │
│ │
│ │
│ EXPANSION TECHNIQUE COMPARISON: │
│ ─────────────────────────────── │
│ │
│ ┌───────────────┬────────────┬────────────┬───────────┐ │
│ │ Technique │ Latency │ Recall↑ │ Precision │ │
│ ├───────────────┼────────────┼────────────┼───────────┤ │
│ │ Synonyms │ Low │ Medium │ May drop │ │
│ │ PRF │ 2x search │ High │ Variable │ │
│ │ Multi-query │ +LLM call │ High │ Good │ │
│ │ HyDE │ +LLM call │ Very high │ Good │ │
│ │ Decomposition │ +LLM call │ High │ Good │ │
│ └───────────────┴────────────┴────────────┴───────────┘ │
│ │
└────────────────────────────────────────────────────────────┘Common questions
Q: Does query expansion always help retrieval?
A: No—expansion increases recall but can hurt precision by introducing irrelevant matches. Synonym expansion is notorious for query drift (e.g., “Python” → “snake”). LLM-based expansion is generally better but adds latency. Always measure impact on your specific use case.
Q: How is HyDE different from regular query expansion?
A: Traditional expansion adds terms to the query but searches in “query space.” HyDE generates a hypothetical answer (full document-like text) and searches in “document space.” Since embeddings of documents cluster differently than query embeddings, HyDE can find documents that query expansion misses.
Q: Should I use query expansion with dense retrieval?
A: Dense retrieval already handles semantic similarity, reducing the need for synonym expansion. However, LLM-based techniques (multi-query, HyDE, decomposition) still help by providing multiple retrieval angles. Combine dense retrieval with multi-query generation for best results.
Q: How do I choose between query expansion techniques?
A: Start with multi-query generation—it’s simple and effective. Use HyDE for knowledge-intensive queries where a hypothetical answer helps. Use decomposition for complex multi-part questions. Avoid classical synonym expansion unless you have domain-specific synonyms.
Related terms
- Dense retrieval — benefits from query expansion
- RAG — pipeline where expansion is applied
- Reranking — follows expansion in pipeline
- Cross-encoder — reranks expanded results
References
Gao et al. (2022), “Precise Zero-Shot Dense Retrieval without Relevance Labels”, ACL 2023. [HyDE paper]
Wang et al. (2023), “Query2Doc: Query Expansion with Large Language Models”, EMNLP. [LLM query expansion]
Carpineto & Romano (2012), “A Survey of Automatic Query Expansion in Information Retrieval”, ACM Computing Surveys. [Classical techniques]
Ma et al. (2023), “Query Rewriting for Retrieval-Augmented Large Language Models”, EMNLP. [RAG query rewriting]