Query Expansion

An automatic, iterative technique that assumes the top-k documents retrieved by an initial query are relevant. The algorithm extracts discriminative terms from these pseudo-relevant documents and appends them to the original query. This is the classic blind relevance feedback approach, often implemented using Rocchio's algorithm for vector space re-weighting.

• Assumption: Top-n results are relevant
• Risk: Query drift if initial results are poor
• Common in: BM25-based search pipelines

A deterministic method that uses a curated lexical database like WordNet to append synonyms, hypernyms, and hyponyms to the query. Unlike statistical methods, this approach relies on manually or automatically constructed controlled vocabularies and is highly interpretable.

• Strength: No training data required
• Limitation: Fails on domain-specific jargon
• Example: Expanding 'automobile' with 'car', 'vehicle', 'sedan'

Leverages dense vector representations to find semantically proximate terms. A seed query is encoded into an embedding space, and the k-nearest neighbors in that space—terms with high cosine similarity—are added to the query. This captures conceptual relationships beyond lexical overlap.

• Uses: Word2Vec, GloVe, or contextual BERT embeddings
• Advantage: Captures analogical relationships
• Integration: Pairs naturally with vector databases and ANN indices

Uses a large language model to generate reformulated queries, hypothetical answers, or related sub-questions. The model is prompted to produce diverse linguistic variations of the original intent. This is the foundation of HyDE (Hypothetical Document Embeddings), where a fake ideal document is generated and its embedding is used for retrieval.

• Prompt: 'Generate 5 alternative phrasings for the question...'
• Benefit: Handles zero-shot, complex intents
• Cost: Higher latency and compute than static methods

Expands the current query using entities and intents extracted from the user's conversational history or session context. This is critical for multi-turn conversational search, where a query like 'its capital' requires resolving the anaphora to a previously mentioned country.

• Mechanism: Coreference resolution + entity linking
• Application: Chatbots and voice assistants
• Key tech: Maintaining a session state with a knowledge graph

A statistical approach that mines historical search logs to find terms that frequently co-occur in sessions leading to successful clicks. Query-URL click graphs are analyzed to build probabilistic term association rules, reflecting real user behavior rather than linguistic theory.

• Data source: Search engine query logs
• Metric: Pointwise Mutual Information (PMI)
• Advantage: Captures domain-specific, emergent vocabulary

A metric measuring the cosine of the angle between two non-zero vectors. It quantifies semantic similarity irrespective of vector magnitude, making it the standard for comparing query and document embeddings.

• Range: -1 (opposite) to 1 (identical)
• Magnitude-invariant: focuses purely on orientation
• Essential for nearest neighbor search in dense vector spaces

A powerful query expansion technique where a language model generates a hypothetical ideal document from a user query. The embedding of this synthetic document is then used to search a vector store for similar real documents.

• Bridges the gap between short queries and detailed documents
• Improves recall by expanding the query into a full document representation
• Particularly effective when queries are terse or ambiguous

A class of algorithms that trade a small amount of accuracy for significant speed improvements when finding similar vectors in high-dimensional spaces. ANN is essential for scalable semantic search over millions or billions of embeddings.

• Sub-linear time complexity vs. brute-force linear scan
• Key algorithms: HNSW, IVF-PQ, LSH
• Enables real-time query expansion retrieval at scale

A metric evaluating the conceptual closeness of two pieces of text based on meaning rather than lexical overlap. Query expansion leverages semantic similarity to find related terms that share the same vector neighborhood.

• Computed via vector distance (cosine, Euclidean, dot product)
• Captures synonyms, paraphrases, and conceptual relationships
• Foundation for dense retrieval and modern information access

A retrieval strategy combining the precision of sparse keyword search (BM25) with the semantic understanding of dense vector search. Query expansion often feeds into both pipelines simultaneously.

• Sparse retrieval: exact term matching for precision
• Dense retrieval: semantic matching for recall
• Fused using Reciprocal Rank Fusion (RRF) for a consolidated ranking

A two-stage retrieval architecture where a fast bi-encoder retrieves candidate documents, and a slower, more accurate cross-encoder jointly processes the query and document to re-rank results. Query expansion improves the initial candidate pool quality.

• Bi-encoder: encodes query and document independently for speed
• Cross-encoder: processes query-document pairs jointly for accuracy
• Critical for maximizing precision in the final result set