Hybrid Search

Maintains two distinct inverted indexes: a dense vector index for semantic similarity and a sparse inverted index for keyword matching. Each document is simultaneously stored as a high-dimensional embedding and a bag-of-words representation. This dual-storage architecture ensures no single retrieval paradigm becomes a bottleneck, allowing the system to handle both conceptual queries and precise identifier lookups without compromise.

A rank aggregation algorithm that merges disparate result lists without requiring score calibration. RRF computes a combined score using the formula 1 / (k + rank) where k is a constant (typically 60). This non-parametric method elegantly solves the scale mismatch between cosine similarity scores and BM25 term frequencies, producing a unified ranking that consistently outperforms linear combination approaches in benchmarks.

Sparse retrieval ensures exact matching for queries containing proper nouns, product codes, error messages, or rare identifiers that dense embeddings often mishandle. For example, searching for ERR-503-XP or ACME-4452 returns zero false positives because BM25's term frequency analysis directly matches the literal string, while vector search might retrieve semantically similar but incorrect error codes.

Dense retrieval captures conceptual relationships beyond lexical overlap, enabling the system to match cost reduction strategies with documents discussing expense optimization techniques. This capability is critical for natural language queries where users describe intent rather than keywords. The embedding space maps synonyms, paraphrases, and related concepts into proximal vector regions, dramatically improving recall for exploratory searches.

Advanced implementations dynamically adjust the fusion ratio based on query characteristics. A query classifier analyzes incoming text to determine whether it is keyword-heavy or natural language, then weights the dense and sparse components accordingly. For instance, a query like What are the main causes of transformer attention drift? receives higher dense weighting, while SKU-99821 inventory triggers stronger sparse emphasis.

Both retrieval paths support pre-filtering on structured metadata fields before similarity computation. Documents can be restricted by date ranges, document type, access permissions, or custom tags prior to vector or keyword search. This ensures that hybrid retrieval respects enterprise security policies and temporal constraints, returning only authorized and relevant documents without post-hoc filtering overhead.

Encodes queries and documents into low-dimensional vector embeddings to perform semantic similarity matching. Unlike sparse methods, dense retrieval understands conceptual relationships—a search for 'CPU cooling solutions' can retrieve documents about 'heat sinks' without keyword overlap. Typically powered by bi-encoder models and indexed using approximate nearest neighbor (ANN) algorithms for sub-linear search times.

Represents text using high-dimensional sparse vectors based on lexical term frequency. Algorithms like BM25 and TF-IDF excel at exact keyword matching and handle rare terms exceptionally well. Sparse retrieval provides the precision backbone in hybrid systems, ensuring that queries for specific product codes or technical jargon return exact matches that dense semantic search might overlook.

The dominant fusion algorithm for combining dense and sparse result sets. RRF scores each document based on its reciprocal rank across multiple ranked lists: score = Σ 1/(k + rank_i). The constant k (typically 60) dampens the impact of high rankings. RRF requires no calibration or score normalization, making it robust when merging heterogeneous relevance signals from vector and keyword search.

A dual-tower neural network that independently encodes queries and documents into separate embedding vectors. This separation allows documents to be pre-indexed offline, enabling efficient maximum inner product search (MIPS) at query time. Bi-encoders trade some accuracy for speed compared to cross-encoders, making them the standard choice for first-stage retrieval in hybrid search pipelines.

A neural model that processes a query-document pair jointly through full self-attention, generating a highly accurate relevance score. Cross-encoders capture fine-grained semantic interactions that bi-encoders miss. In hybrid search architectures, they serve as a second-stage re-ranker applied to the fused candidate set, significantly improving precision at the cost of higher computational overhead.

A retrieval paradigm that encodes queries and documents into token-level multi-vector representations and computes relevance via a MaxSim operation. Unlike single-vector bi-encoders, ColBERT preserves per-token embeddings, enabling fine-grained lexical and semantic matching. It bridges the gap between the efficiency of dense retrieval and the precision of cross-encoders without requiring joint computation.