RAG Query Expansion: How Synthetic Wording Finds Evidence Beyond Exact Terms

Synthetic query expansion often improves RAG recall by adding plausible paraphrases and domain terms that bridge vocabulary gaps between questions and stored evidence.

A family member may ask for “the boiler inspection,” while the scanned invoice says “annual hydronic service” and never uses the household phrase. Exact lexical search misses that connection. Query expansion generates alternate wording or a hypothetical answer, searches each representation, and merges the candidates before reranking, giving relevant evidence more than one route into the context window.

One Question Can Represent Several Search Intents

A short household query may omit the formal product name, date, location, or terminology used in the source. Expansion makes those latent interpretations explicit by producing related phrases, subqueries, or a synthetic passage. Sparse retrieval gains new matching terms, while dense retrieval receives additional semantic viewpoints.

The Query2doc method method prompts a language model to generate a pseudo-document and appends it to the original query. Its experiments show that the added text can improve both sparse and dense retrieval without retraining the underlying retriever.

The mechanism does not make the query more truthful. It broadens the candidate-generation surface, after which fusion and reranking must decide which candidates still answer the original question rather than the generated elaboration. This distinction remains visible during later household testing.

Synthetic Passages Bridge the Query-Document Vocabulary Gap

A hypothetical document can resemble the style and vocabulary of likely evidence even when its facts are invented. The system embeds that synthetic passage and retrieves real documents near it, using generation as a semantic bridge rather than as evidence presented to the user.

hypothetical document embeddings creates hypothetical documents before embedding them for zero-shot dense retrieval. The important separation is that generated text guides retrieval, while the returned corpus documents remain the only admissible evidence for the final answer.

Multiple expansions can cover abbreviations, nicknames, translations, and different levels of technical detail. Candidate union increases recall, but deduplication and reranking are necessary because ten rewrites can otherwise crowd the list with near-identical or weakly related chunks.

Expansion Can Drift Toward a Plausible but Wrong Topic

A generator may resolve an ambiguous name incorrectly, introduce a product the household does not own, or invent a date. Those terms can retrieve internally consistent documents about the wrong entity, creating more confident noise than the original sparse query.

A controlled study of complementary query rewriting finds that combinations can improve coverage on some datasets while hurting an ambiguous-question benchmark. That result shows why expansion should be gated by query type and evaluated against a strong baseline.

The failure boundary is any query whose central entity or intent is uncertain. Preserve the original query, label synthetic text as a retrieval aid, cap its influence during fusion, and abstain when expansions disagree on the evidence set.

Featured

ZimaCube 2 Personal Cloud Home NAS

ZimaCube2

$799.00 - $2,499.00

Buy Now

Run an Expansion Contribution Test

Create thirty real household questions containing abbreviations, informal names, translations, and formal terms absent from the query. Record the verified supporting chunks, then run original-only retrieval and the proposed expansion route with the same index, top-k budget, and reranker.

Track recall and citation coverage using the evaluation method in abbreviation retrieval. For each added relevant chunk, store which rewrite found it; for every new distractor, record the term that caused drift and whether reranking removed it.

Enable expansion only where it raises held-out recall without unacceptable precision loss or latency. If gains come merely from a larger candidate budget, compare against an original-query run with the same budget before attributing the improvement to synthetic wording.