Query Fusion

Reciprocal Rank Fusion (RRF) is a rank aggregation method that combines multiple ranked lists without requiring relevance scores. It assigns a new score to each document based on its reciprocal rank in each list, favoring documents that appear consistently high across different result sets.

• Mechanism: The score for a document d is calculated as: score(d) = Σ (1 / (k + rank_i(d))), where rank_i(d) is the document's rank in list i, and k is a constant (typically 60) to dampen the effect of high ranks.
• Advantage: It is highly effective for fusing results from heterogeneous retrievers (e.g., combining dense retrieval with BM25) because it is score-agnostic and robust to different score distributions.
• Use Case: The primary method for late-stage fusion in hybrid retrieval systems to merge lexical and semantic search results.

Score-based fusion techniques normalize and combine the relevance scores from different retrieval runs. The two most common algorithms are CombSUM and CombMNZ.

• CombSUM: Sums the normalized scores a document receives across all result lists. It rewards documents that have high relevance in multiple lists.
• CombMNZ (Combined by Multiple of Non-Zero): Multiplies the CombSUM score by the number of lists that retrieved the document. This strongly boosts documents retrieved by many sources, acting as a consensus signal.
• Critical Step: Requires careful score normalization (e.g., Min-Max, Z-score) before combination, as scores from different models (e.g., a cross-encoder vs. a bi-encoder) are not directly comparable.
• Application: Often used after an initial query expansion round to fuse results from the original and expanded queries.

Query-variant fusion executes multiple representations of the same user query and fuses their results. This strategy hedges against the ambiguity or inadequacy of any single query formulation.

• Common Variants Generated Via:
  • Query Reformulation: Using an LLM to rephrase the query for clarity or specificity.
  • Query Expansion: Adding synonyms or related terms via techniques like Pseudo-Relevance Feedback (PRF).
  • Intent Disambiguation: Running parallel retrievals for different possible interpretations of an ambiguous query.
• Process: Each variant retrieves a candidate set, which are then merged using RRF or score-based fusion.
• Benefit: Dramatically improves recall by covering a broader semantic space, which is crucial for complex or underspecified queries.

Multi-representation fusion leverages different representations of the same documents for a single query. Instead of multiple queries, it uses multiple retrieval models that interpret document content differently.

• Classic Implementation: Hybrid Search, which fuses results from:
  • Sparse Retrieval (e.g., BM25): Excels at exact keyword and lexical matching.
  • Dense Retrieval: Excels at capturing semantic similarity via query embedding and document embeddings.
• Advanced Implementation: Using multiple dense encoders (e.g., trained on different objectives or data) to create diverse embedding spaces for the same corpus.
• Outcome: Achieves a balance of high precision (from semantic matches) and high recall (from lexical matches), making it the backbone of modern Retrieval-Augmented Generation (RAG) systems.

Using a cross-encoder for reranking can be viewed as a sophisticated fusion technique. It doesn't combine lists but uses a powerful model to re-evaluate a fused candidate set, effectively performing a final, learned aggregation.

• Process: An initial broad retrieval (often itself a fusion of methods like hybrid search) produces a candidate set of 100-200 documents. A computationally intensive cross-encoder model then scores each (query, document) pair jointly, producing a new, high-precision ranked list.
• Fusion Aspect: The cross-encoder implicitly fuses all features of the query and document—lexical, semantic, and syntactic—into a single relevance judgment. It acts as the ultimate arbiter, correcting errors from the first-stage retrievers.
• System Role: This is a critical component in multi-stage retrieval pipelines for achieving production-grade accuracy, directly feeding the most relevant context into an LLM.

Learned fusion employs machine learning models, specifically Learning to Rank (LTR) algorithms, to optimally combine multiple relevance signals into a single ranking. This moves beyond heuristic formulas like RRF.

• Input Features: The model is trained on features such as:
  • Scores from individual retrievers (BM25, dense retrieval).
  • Rank positions from different lists.
  • Document metadata (popularity, freshness).
  • Query characteristics (length, detected intent).
• Algorithms: Can use pointwise, pairwise, or listwise LTR methods (e.g., LambdaMART).
• Advantage: Can learn complex, non-linear relationships between signals that are optimal for a specific domain or dataset, outperforming static heuristics.
• Challenge: Requires substantial labeled training data (<query, document, relevance_score> tuples) and ongoing maintenance as data distributions shift.

A retrieval technique that augments an original user query with additional relevant terms or phrases to improve recall by matching a broader set of documents. Common methods include:

• Pseudo-Relevance Feedback (PRF): Automatically extracts terms from the top-ranked results of an initial search.
• Synonym addition: Uses lexical databases or embedding neighborhoods to add conceptually similar words.
• Controlled vocabulary mapping: Links query terms to domain-specific thesauri or ontologies. This technique addresses the vocabulary mismatch problem but can reduce precision if expansion terms are not carefully selected.

The process of altering a user's original query to better align with the underlying information need. Unlike expansion, reformulation often rewrites or corrects the query. Key approaches include:

• Spelling and grammar correction.
• Query simplification (e.g., removing stop words, reducing complexity).
• Disambiguation (e.g., clarifying 'Apple' as the company vs. the fruit).
• Conversational reformulation: Resolving co-references (e.g., changing 'it' to the previously mentioned entity) in multi-turn dialogues. This process is critical for handling ambiguous, noisy, or conversational user inputs.

A neural search paradigm where queries and documents are encoded into dense vector embeddings (typically 768 or 1024 dimensions). Relevance is computed via similarity measures like cosine similarity in this high-dimensional space.

• Encoders: Models like BERT, Sentence-BERT, or Contriever are used to generate embeddings.
• Advantage: Captures semantic meaning beyond keyword matching (e.g., 'car' matches 'automobile').
• Challenge: Requires significant compute for embedding generation and efficient nearest neighbor search, often powered by a vector database. In a hybrid retrieval system, dense retrieval results are a primary candidate for fusion with sparse (keyword-based) results.

A precision-focused technique where a computationally intensive model scores the relevance between a query and each candidate document individually. It is typically applied after an initial, fast retrieval step (like BM25 or dense retrieval).

• Mechanism: The query and a document text are concatenated and fed into a transformer model (e.g., a fine-tuned BERT) that outputs a relevance score.
• Role in Fusion: The scores from a cross-encoder are often used as a high-quality signal within a fusion algorithm (like Reciprocal Rank Fusion) to boost the rank of documents it deems highly relevant, significantly improving final precision.

The task of categorizing a user's search query into a predefined intent type to guide the retrieval and ranking strategy. Common intent taxonomies include:

• Informational: Seeking knowledge (e.g., 'What is query fusion?').
• Navigational: Looking for a specific website or page (e.g., 'Inferensys blog').
• Transactional: Aiming to perform a transaction (e.g., 'buy GPU').
• Commercial Investigation: Comparing products before a purchase. Intent signals can directly influence fusion strategies; for example, a navigational intent might prioritize lexical matches from a sparse retriever, while an informational intent might weight semantic matches from a dense retriever more heavily.

The overarching architecture that combines multiple retrieval methods—typically sparse retrieval (e.g., BM25) and dense retrieval—to balance the strengths of each. Sparse retrieval excels at exact keyword matching and term importance, while dense retrieval captures semantic similarity.

• Query fusion is a core technique within hybrid retrieval, used to merge the ranked lists from these disparate retrievers.
• Other combination methods include learning-to-rank models that use scores from both systems as features. The goal is to achieve higher recall and precision than any single retriever could provide alone.