Query Rewriting

Augments the original query with additional terms to improve recall. This can be achieved through synonym injection using a thesaurus or word embeddings, or by adding hypernyms and hyponyms to broaden or narrow the search scope. For example, a query for 'laptop' might be expanded to include 'notebook computer' and 'ultrabook'.

• Classic approach: Pseudo-Relevance Feedback (PRF) extracts terms from top-k initial results
• Modern approach: Generative models like LLMs produce expanded query variants
• Risk: Over-expansion can introduce noise and reduce precision

Breaks a complex, multi-faceted question into a sequence of simpler sub-queries that can be answered independently. This is essential for multi-hop reasoning where the answer to one sub-question serves as context for the next.

• Sequential decomposition: 'What is the capital of the country where the Amazon rainforest is located?' becomes 'Where is the Amazon rainforest?' → 'Brazil' → 'What is the capital of Brazil?'
• Parallel decomposition: Independent facets are retrieved simultaneously and answers are aggregated
• Tool use: LLMs with chain-of-thought prompting generate structured sub-query plans

A technique where a language model generates a hypothetical ideal document that would answer the query, then uses the embedding of that synthetic document for retrieval rather than the query embedding itself. This bridges the vocabulary gap between short queries and verbose documents.

• A query like 'how to fix a leaky faucet' might generate a hypothetical DIY guide paragraph
• The dense embedding of that generated text is used for ANN search
• Particularly effective when queries and documents have different linguistic styles or lengths

Generates a higher-level, more abstract 'step-back' question that captures the general concept before addressing the specific query. This improves retrieval by first grounding the system in broad, relevant context.

• Example: For 'What happens to the pressure of an ideal gas if volume doubles at constant temperature?', the step-back question is 'What is the ideal gas law?'
• Retrieves foundational knowledge first, then applies it to the specific case
• Reduces hallucination by ensuring the model has access to core principles before reasoning

Uses a language model to generate a pseudo-document from the query, then concatenates that generated text with the original query to form an expanded search string. Unlike HyDE which uses only the synthetic embedding, Query2Doc enriches the query with additional lexical terms for sparse retrieval.

• The pseudo-document adds contextually relevant keywords
• Benefits both dense and sparse retrieval pipelines
• Can be combined with BM25 to inject learned vocabulary into keyword search

Generates multiple semantically diverse reformulations of the same user query, retrieves documents for each variant independently, and then applies Reciprocal Rank Fusion (RRF) to merge the results into a single ranked list.

• Each reformulation captures a different perspective or phrasing of the same intent
• Improves recall by covering multiple ways the answer might be expressed in the corpus
• The fusion step automatically up-weights documents that appear consistently across multiple retrieval passes

An algorithm that merges multiple ranked result lists without requiring score calibration. It assigns a score based on the reciprocal of each document's rank position.

• Formula: score(d) = Σ 1/(k + rank_i(d)) where k is a constant (typically 60)
• Key benefit: Outperforms linear combination when retrieval sources produce uncalibrated scores
• Use case: Merging results from a rewritten query's dense and sparse retrieval passes

The task of categorizing a raw query into a predefined taxonomy of user goals before rewriting or retrieval occurs. This classification gates which downstream rewriting strategies are applied.

• Example intents: factual lookup, comparative analysis, procedural how-to, transactional
• Impact on rewriting: A factual query may trigger expansion, while a procedural query triggers decomposition
• Implementation: Typically a fine-tuned classifier or few-shot prompted LLM

A blind relevance feedback technique that assumes the top-k documents from an initial retrieval are relevant, extracts discriminative terms from them, and appends those terms to the original query for a second retrieval pass.

• Contrast with rewriting: PRF is post-retrieval expansion driven by corpus statistics, not semantic understanding
• Risk: Query drift if top-k documents are not actually relevant
• Modern variant: Using LLMs to summarize top-k documents into a refined query

A cascading pipeline architecture where rewritten queries first hit a fast, high-recall candidate generation stage, followed by progressively more expensive and precise re-ranking stages.

• Stage 1: Sparse (BM25) or dense (ANN) retrieval over millions of documents
• Stage 2: Cross-encoder re-ranking of top-100 candidates
• Query rewriting's role: Ensures the first stage has the lexical and semantic coverage to surface relevant candidates

A strategy that executes multiple heterogeneous retrieval systems in parallel—dense vector search, sparse keyword matching, and metadata filtering—and fuses their results. Query rewriting often generates the distinct formulations fed to each system.

• Example: A rewritten query sends a keyword-expanded version to BM25 and a semantically abstracted version to the vector index
• Fusion methods: RRF, weighted sum, or learning-to-rank combiner
• Goal: Maximize recall by covering each system's blind spots

A precision-focused re-ranking methodology where a transformer processes the concatenated query and candidate document jointly to produce a fine-grained relevance score. It validates whether rewritten queries actually retrieved relevant content.

• Latency cost: Orders of magnitude slower than bi-encoder scoring
• Placement: Applied to top-50 or top-100 candidates from first-pass retrieval
• Synergy with rewriting: A well-rewritten query surfaces better candidates, reducing the burden on the expensive cross-encoder stage