Hybrid Search

Sparse retrieval, exemplified by algorithms like BM25 (Best Matching 25), operates on exact keyword matching. It uses an inverted index to map terms to documents.

• Strengths: Excellent for precise keyword queries, proper nouns, and acronyms. Highly interpretable and computationally efficient.
• Weaknesses: Fails with synonyms, paraphrasing, or semantic intent (e.g., querying 'automobile' won't match documents containing only 'car').
• Core Mechanism: Scores documents based on term frequency (TF) and inverse document frequency (IDF), penalizing very long documents.

Dense retrieval uses neural embedding models (e.g., Sentence-BERT, OpenAI embeddings) to encode queries and documents into high-dimensional vector embeddings.

• Strengths: Captures semantic meaning, enabling matches based on conceptual similarity, not just lexical overlap. Robust to vocabulary mismatch.
• Weaknesses: Can miss critical exact keyword matches; performance depends heavily on the embedding model's domain alignment.
• Core Mechanism: Similarity is computed via cosine similarity or dot product in a shared vector space, facilitated by a dense vector index like HNSW or IVF.

The core of hybrid search is combining scores from sparse and dense retrievers. Common fusion methods include:

• Weighted Sum: final_score = (alpha * sparse_score) + ((1 - alpha) * dense_score). The alpha hyperparameter controls the blend.
• Reciprocal Rank Fusion (RRF): Combines result rankings without needing normalized scores, often more robust.
• Learned Reranking: A cross-encoder model (e.g., BGE-reranker) re-evaluates the top-k combined results for precise final ordering, adding computational cost but significantly improving accuracy.

Hybrid search systems integrate metadata filtering as a pre-retrieval or post-retrieval step to enforce hard business logic.

• Pre-Filtering: Scopes the search corpus before semantic/keyword search (e.g., WHERE date > 2024 AND department = 'Engineering'). Ensures compliance but can exclude relevant results if the filter is too strict.
• Post-Filtering: Applies filters after retrieving a broad set of results. Preserves recall but may leave fewer final results.
• Use Case: Essential for enterprise RAG, allowing search within specific date ranges, document sources, access control groups, or other structured attributes.

To optimize input for both retrieval paths, queries are often processed before search execution.

• For Sparse Retrieval: Query expansion adds synonyms or related terms from a knowledge graph or LLM to improve recall. Spell correction is also critical.
• For Dense Retrieval: The raw query may be rewritten by an LLM for clarity or to generate multiple hypothetical document embeddings.
• Hybrid Approach: A system might generate a keyword query for the sparse retriever and a semantically refined query for the dense retriever from the same user input.

Production hybrid search requires co-located or synchronized indices.

• Unified Databases: Modern vector databases (e.g., Weaviate, Qdrant, Pinecone) natively support hybrid search by maintaining combined inverted and vector indices.
• Dual-Path Orchestration: A retrieval service may query a separate Elasticsearch (sparse) cluster and a Faiss (dense) index, then fuse results in the application layer.
• Incremental Indexing: Both indices must support real-time or near-real-time updates to reflect changing data, often via a log-based change data capture (CDC) pipeline.

Sparse retrieval is a family of information retrieval methods that rely on exact lexical matches between query terms and documents. The most prominent algorithm is BM25 (Best Matching 25), a probabilistic function that scores documents based on:

• Term frequency (TF): How often a query term appears in a document.
• Inverse document frequency (IDF): How rare the term is across the entire corpus.
• Document length normalization: Penalizes very long documents that may accumulate matches by sheer size. It excels at finding documents containing specific keywords, names, or acronyms but fails with synonyms or paraphrases. In hybrid search, it provides the high-precision, keyword-matching component.

Dense retrieval represents queries and documents as high-dimensional vector embeddings generated by a neural network (e.g., Sentence-BERT). Relevance is measured by cosine similarity between vectors. Key aspects include:

• Semantic understanding: Captures conceptual meaning, enabling matches for synonyms and related ideas.
• Embedding model: The quality of retrieval is directly tied to the model used to create the embeddings.
• Approximate Nearest Neighbor (ANN) search: Algorithms like HNSW or IVF enable fast search over billions of vectors. It forms the semantic, recall-oriented pillar of hybrid search but can be sensitive to vocabulary mismatch.

Score fusion is the mechanism that combines ranked lists from sparse and dense retrievers into a single, unified result set. Two primary methods are:

• Reciprocal Rank Fusion (RRF): A robust, parameter-free method that combines rankings based on the reciprocal of the rank of each document in each list. It is highly effective for balancing disparate score distributions.
• Weighted Score Fusion: Manually or learned weighted linear combination of normalized scores from each retriever (e.g., final_score = α * bm25_score + (1-α) * cosine_similarity). This stage is critical for determining the final ranking and trade-off between precision and recall in the hybrid output.

Reranking is a subsequent, computationally intensive stage where a more powerful model (a cross-encoder) re-evaluates a small candidate set (e.g., top 100 results from hybrid search).

• Cross-Encoder: A model that takes a query and a single document as a concatenated input, allowing deep, attention-based interaction for a highly accurate relevance score.
• Precision Optimization: It does not search the full corpus but refines the final ranking of the hybrid search's shortlist, pushing the most relevant document to the top.
• Trade-off: Adds latency but significantly improves result quality for critical applications.

ColBERT (Contextualized Late Interaction over BERT) is a neural retrieval model that represents a middle ground between sparse and dense retrieval. Its architecture involves:

• Per-token embeddings: Encodes every token in both the query and document into a contextualized embedding.
• Late interaction: Scores relevance by computing the MaxSim operation—for each query token, finding its maximum similarity with any document token, then summing these maxima.
• Efficiency: Document embeddings can be pre-computed and indexed. It provides richer matching than single-vector dense retrieval (capturing partial matches like 'CPU' to 'processor') and is often used as a powerful component in multi-stage hybrid systems.

Metadata filtering is a pre- or post-retrieval step that applies hard constraints based on document attributes, acting as a boolean filter on the candidate pool. Common uses in hybrid search include:

• Pre-filtering: Restricting the search corpus to documents with source = 'internal_wiki' and date > 2023 before executing vector or keyword search, drastically improving performance.
• Post-filtering: Applying filters like language = 'en' after retrieval to ensure result compliance.
• Hybrid Integration: Often implemented as a filtered search where the ANN graph traversal or BM25 search only considers documents passing the metadata predicates. It is essential for production systems requiring access control or domain-specific scoping.