Generative Query Expansion

Unlike rule-based methods, this technique uses a prompt to instruct a sequence-to-sequence model. The model reads the original query and generates a list of expansion terms, synonyms, or even full alternative queries. This allows for zero-shot expansion where the model invents terms not present in its training data, guided purely by the prompt's instructions.

The foundational architecture often relies on encoder-decoder models like T5 (Text-to-Text Transfer Transformer). The encoder processes the input query, and the decoder autoregressively generates expansion tokens. This is distinct from embedding-based expansion because the model actively generates new text strings rather than just retrieving similar vectors.

A core capability is generating syntactically diverse paraphrases that preserve semantic intent. For a query like 'cheap flights to NYC', the model might generate 'affordable airfare to New York' or 'budget tickets JFK'. This bridges the vocabulary gap between user jargon and formal document terminology without manual synonym curation.

A critical engineering challenge is semantic drift. The generative model may produce fluent but irrelevant terms. Mitigation strategies include:

• Constrained Decoding: Forcing the model to only generate tokens from a valid entity list.
• Verification Reranking: Using a cross-encoder to score the relevance of generated terms against the original query before adding them to the retrieval index.

Generative expansion introduces inference latency at query time, as it requires a forward pass through a large language model. This is computationally more expensive than static dictionary lookups. Production systems often cache frequent query expansions or use model distillation to compress the generator into a smaller, faster student model suitable for high-throughput search pipelines.

The generated terms can be injected into multiple retrieval stages:

• Sparse (BM25): Generated keywords are added directly to the query string with appropriate boosting weights.
• Dense (Vector Search): The expanded query text is encoded into a dense embedding, creating a richer semantic vector that captures the broader intent for approximate nearest neighbor (ANN) search.

Adds words with identical or highly similar meanings to the original query terms to increase recall.

• Example: Expanding "car" with "automobile" and "vehicle"
• Often uses curated thesauri or WordNet for structured synonym sets
• Differs from generative expansion by relying on static lexical resources rather than dynamic model inference

An automatic technique that assumes the top-k documents from an initial retrieval are relevant and extracts expansion terms from them.

• Classic implementation uses the Rocchio Algorithm to reformulate query vectors
• Contrasts with generative expansion: PRF is post-retrieval and corpus-dependent
• Risk of query drift if top documents are not truly relevant

Uses a model to generate alternative phrasings of the original query that convey the same meaning.

• Example: "How to fix a bike tire" → "Bicycle tire repair instructions"
• Back-Translation Expansion is a specific technique that translates to an intermediate language and back
• Generative Query Expansion is a superset that includes paraphrase generation plus term-level augmentation

Uses deep language models like BERT to generate expansion terms that are semantically appropriate for the query's specific context.

• Unlike static Word Embedding Expansion (Word2Vec), disambiguates polysemous words
• Example: "apple" in a tech context expands to "iPhone" and "MacBook", not "fruit"
• Generative expansion with T5 or GPT extends this by producing multi-word phrases and full alternative queries

Enriches a query by traversing an entity's relationships in a knowledge graph to add related entities and attributes.

• Example: Query for "Tesla" expands with "Elon Musk", "electric vehicles", "Model 3"
• Relies on entity linking to first ground the query to a graph node
• Generative expansion can complement this by synthesizing natural language expansions from graph triples

A document expansion technique that uses a sequence-to-sequence model to generate potential queries a document might answer, appended to the document's index.

• Inverse of query expansion: enriches the document side rather than the query side
• Trained on query-document pairs to predict queries from document text
• Often used alongside generative query expansion for dual-sided retrieval improvement