Query2Doc

Query2doc Query-Expansion Dense-Retrieval Neural-Ir Needs-Review

What it is

Query2Doc (Wang et al., 2023) improves retrieval by expanding the query with a few-shot LLM-generated pseudo-document rather than replacing it. The expanded query (original query + generated passage) is then used with BM25 or a dense encoder. Unlike HyDE, which embeds only the generated text, Query2Doc concatenates original and generated text, preserving the exact query terms while adding semantic context.

[illustrate: Query + few-shot prompt → LLM → pseudo-document → concatenate with original query → BM25 or dense retrieval]

How it works

  1. Few-shot prompt:

    Q: {example_query_1}  A: {example_answer_1}
Q: {example_query_2}  A: {example_answer_2}
Q: {query}            A:

  2. Generation:

    • LLM generates a short pseudo-document (answer-like passage)

  3. Query expansion:

    • For BM25: concatenate query + pseudo-document as the query string
    • For dense: encode the concatenated text as the query vector

  4. Retrieval:

    • Standard BM25 or ANN retrieval with the expanded query

Variants and history

Query2Doc (2023) showed improvements on MSMARCO and TREC DL for both sparse and dense retrieval. Compared to HyDE: Query2Doc retains the original query terms (important for entity-heavy or code queries), while HyDE discards them. Step-Back Prompting and Chain-of-Thought Retrieval are related expansions. In practice, HyDE and Query2Doc often have complementary strengths.

When to use it

Use Query2Doc when:

  • LLM is available for query-time inference
  • Original query terms must be preserved (entities, technical terms)
  • Both sparse and dense retrieval are in use (Query2Doc helps both)
  • Few-shot examples for the domain are available

See also