Query Expansion with Language Models

A zero-shot paradigm where the LLM generates a hypothetical ideal document that would perfectly answer the query, rather than expanding the query itself. The dense vector embedding of this synthetic text is then used for similarity search against the real corpus.

• Mechanism: Query → LLM generates fake answer → Embed fake answer → Vector search
• Key Insight: Grounding retrieval in the dense representation of an answer, not the question, aligns the search space with the document space.
• Contrast: Differs from traditional PRF by generating synthetic text from the model's parametric knowledge rather than extracting terms from an initial retrieval pass.

The LLM is prompted to act as a search expert, generating a list of highly relevant keywords, synonyms, and related entities based on the original query's latent intent. This bridges the lexical gap where users and authors use different terminology for the same concept.

• Example: Query 'how to fix a leaky faucet' → Expanded terms: 'drip repair', 'cartridge replacement', 'O-ring', 'valve seat', 'plumbing'
• Advantage: Captures domain-specific jargon and multi-word expressions that simple thesaurus-based expansion misses.
• Risk: Unconstrained generation can introduce topic drift if the model hallucinates unrelated concepts.

For complex analytical queries, the LLM breaks the question into a structured plan of sub-queries before expansion. Each sub-query targets a specific facet, and their expansions are federated to retrieve comprehensive evidence.

• Process: 'Compare Q3 revenue of Company A and B' → 1. 'Company A Q3 revenue report' 2. 'Company B Q3 earnings call transcript' 3. 'Sector performance analysis'
• Benefit: Prevents the 'information bottleneck' where a single expanded query cannot capture all dimensions of a multi-hop question.
• Implementation: Often paired with tool-calling to execute each sub-query against different indexes or APIs.

The LLM reformulates the user's terse or malformed query into a well-formed, declarative statement that mirrors the prose style of the target documents. This shifts the query from a 'search string' to a 'mini-document'.

• Input: 'ml bias mitigation techniques'
• Output: 'A survey of established methodologies for identifying and mitigating statistical bias in machine learning training data and model outputs.'
• Mechanism: The dense embedding of the re-written prose has higher cosine similarity with academic abstracts and technical documentation than the original keyword string.
• Contrast with HyDE: Rewriting stays faithful to the query's scope; HyDE fabricates a speculative answer.

The LLM generates both an inclusive expansion (broadening recall) and a negative constraint (tightening precision) to define the semantic boundary of the query. Negative terms are used for filtering or down-weighting.

• Query: 'Java memory management'
• Inclusive: 'JVM heap allocation', 'garbage collection algorithms', 'GC tuning'
• Negative: 'Java island', 'JavaScript', 'coffee'
• Application: Critical for disambiguating polysemous terms where simple expansion would amplify noise. The negative set acts as a semantic guardrail.

The LLM infers implicit temporal or authority constraints from the query and injects explicit scoping directives into the expanded retrieval call. This moves beyond term expansion to metadata-aware query construction.

• Query: 'latest react best practices'
• Inferred Scoping: date >= 2024, source_type IN (official_docs, verified_expert)
• Query: 'FDA guidance on AI diagnostics'
• Inferred Scoping: domain = fda.gov, document_type = guidance
• Impact: Dramatically reduces retrieval noise by combining semantic expansion with structured filtering, a hybrid approach that pure vector search cannot achieve alone.

Transforms vague product queries into precise retrieval operations. A search for 'lightweight jacket for drizzle' is expanded with terms like 'water-resistant', 'packable', 'windbreaker', and 'softshell'.

• Matches products whose descriptions use technical fabric terminology
• Bridges the gap between buyer vocabulary and catalog taxonomy
• Reduces null results by 40-60% for long-tail queries
• Generated hypothetical answers describe the ideal product attributes before vector search

Expands attorney queries to surface relevant case law despite terminological mismatches. A query about 'slip and fall liability' is augmented with 'premises liability', 'duty of care', 'negligent maintenance', and 'tort law'.

• LLMs generate hypothetical judicial opinions as expansion documents
• Captures jurisdiction-specific legal terminology automatically
• Improves recall in Boolean keyword systems without manual synonym lists
• Preserves precision by anchoring expansions to the original legal concept

Enables clinicians to find relevant studies using colloquial symptom descriptions. 'Why do some diabetics get tingling in their feet?' expands to 'diabetic peripheral neuropathy pathophysiology', 'paresthesia diabetes mellitus', and 'small fiber neuropathy HbA1c correlation'.

• Maps patient-friendly language to MeSH (Medical Subject Headings) terminology
• Generated hypothetical abstracts serve as dense retrieval queries
• Critical for systematic reviews and evidence-based medicine workflows
• Reduces reliance on professional medical librarian intermediaries

Decodes user-reported error messages into root-cause documentation. A query like 'screen goes black after login' is expanded with 'display driver timeout', 'explorer.exe crash', 'black screen of death Windows 11', and 'session initialization failure'.

• LLMs generate a hypothetical troubleshooting article as the expansion vector
• Bridges end-user symptom language and engineering root-cause terminology
• Dramatically improves self-service deflection rates
• Adapts to new error patterns without manual taxonomy updates

Connects interdisciplinary queries to relevant but terminologically distant fields. A search for 'how memory works in computers vs humans' expands to 'von Neumann architecture memory hierarchy', 'long-term potentiation synaptic plasticity', and 'working memory Baddeley model'.

• Surfaces papers from computer architecture and cognitive neuroscience simultaneously
• Hypothetical document generation creates a synthetic interdisciplinary abstract
• Enables serendipitous discovery across siloed academic databases
• Overcomes the disciplinary vocabulary barrier that limits keyword search

Translates investor questions into precise financial metrics and filing terminology. 'Companies that might do well if interest rates drop' expands to 'interest rate sensitivity duration analysis', 'rate cut beneficiaries REITs utilities', 'discounted cash flow lower WACC', and 'refinancing tailwind balance sheet'.

• Maps conversational investment theses to quantifiable financial indicators
• Generated hypothetical analyst reports guide dense retrieval over SEC filings
• Captures sector-specific jargon and acronym expansions automatically
• Enables non-professional investors to access institutional-grade research tools

A classic query expansion technique that assumes the top-k documents from an initial retrieval are relevant. Key terms are extracted from these documents and added to the original query for a second, improved retrieval pass. Unlike LLM-based expansion, PRF relies on statistical term co-occurrence rather than semantic understanding. The technique is computationally lightweight but can suffer from topic drift if the initial results contain noise.

A specific LLM-driven expansion technique where a model generates a hypothetical ideal document from a query. The dense embedding of this generated text is then used to perform a vector similarity search against a real document corpus. HyDE is particularly effective for asymmetric search where queries are short and documents are long, as it bridges the representation gap by creating a document-length proxy.

The technique of reformulating a user's original query into a more effective version for the retrieval system. Unlike expansion, rewriting may replace terms entirely rather than adding to them. Common operations include:

• Correcting spelling errors
• Resolving pronouns and anaphora
• Translating colloquial language to formal terminology
• Decomposing complex questions into atomic sub-queries

A modern retrieval paradigm that encodes queries and documents into dense, fixed-length vector embeddings using neural networks. Query expansion with LLMs often feeds directly into dense retrieval pipelines, where the expanded query is embedded and matched against document vectors via approximate nearest neighbor (ANN) search. This enables semantic matching that goes beyond exact keyword overlap.

An integrated retrieval strategy that combines results from dense vector search and sparse keyword search like BM25. Query expansion techniques can feed both paths simultaneously:

• Expanded keywords boost BM25 recall
• Generated context enriches the dense embedding
• Final results are merged via reciprocal rank fusion or learned scoring This dual approach ensures both semantic understanding and exact term matching.

The process of breaking down a complex, multi-faceted query into a set of simpler, atomic sub-queries that can be independently resolved. LLMs are particularly adept at this task, generating sub-questions that target different aspects of the original query. The answers are later synthesized into a composite response. This is a form of multi-hop reasoning that complements expansion techniques.