Hybrid Search

Encodes queries and documents into high-dimensional embedding vectors using a transformer model. Retrieval is performed by finding the k-nearest neighbors in this embedding space using cosine similarity or Euclidean distance. This captures semantic meaning, paraphrasing, and conceptual relationships that keyword matching would miss.

• Excels at understanding user intent behind vague or conversational queries
• Handles synonyms and cross-lingual matching natively
• Requires an approximate nearest neighbor (ANN) index for production-scale latency

Relies on traditional inverted index structures and algorithms like BM25 or TF-IDF to match exact query terms against documents. It scores documents based on term frequency and inverse document frequency, providing high precision for queries containing rare, specific identifiers.

• Unmatched for exact string matching: part numbers, error codes, legal citations
• Computationally cheap and highly interpretable
• Fails on vocabulary mismatch where the query and document use different words for the same concept

The dominant algorithm for merging separate dense and sparse result lists into a single, unified ranking. RRF calculates a score for each document based on its reciprocal rank across both lists: score = Σ 1 / (k + rank), where k is a constant (typically 60).

• Does not require score calibration between heterogeneous retrieval systems
• Rewards documents that appear high in both lists, surfacing consensus results
• Outperforms linear combination score fusion in most information retrieval benchmarks

A pre-retrieval routing layer that analyzes the incoming query to determine the optimal weighting between dense and sparse retrieval paths. A query for a specific SKU triggers a heavier sparse weight, while a conceptual how-to question biases toward dense.

• Prevents dilution of precise queries with semantically related but incorrect results
• Can be implemented as a lightweight classifier model or rule-based heuristic
• Critical for domains with mixed query types, such as e-commerce or technical documentation

The quality of dense retrieval is entirely dependent on the underlying embedding model. Production systems often use fine-tuned variants of BGE, E5, or GTE models. Key considerations include the model's maximum token length, its performance on the MTEB (Massive Text Embedding Benchmark), and its alignment with the target domain's vocabulary.

• Asymmetric models use separate encoders for queries and documents
• Matryoshka embeddings allow truncation of vector dimensions without catastrophic loss
• Model choice dictates the dimensionality of the vector index and associated storage costs

Hybrid retrieval often produces a larger candidate set than a single method. A computationally more expensive cross-encoder model is then applied to this fused candidate set to perform a final, high-precision re-ranking. Unlike the bi-encoder used for initial retrieval, the cross-encoder processes the query and document jointly.

• Dramatically improves final precision (e.g., NDCG@10)
• Too slow to run over the entire corpus, hence the multi-stage architecture
• Models like Cohere Rerank or BGE-Reranker are purpose-built for this task

The dominant algorithm for merging results from sparse and dense retrieval pipelines. RRF assigns a score to each document based on its reciprocal rank across multiple ranked lists: score = Σ 1/(k + rank_i), where k is a constant (typically 60). This elegantly handles disparate score distributions without calibration. Key properties: non-parametric, score-agnostic, and resistant to outliers. Unlike weighted linear combination, RRF requires no score normalization step.

The traditional half of hybrid search, relying on exact term matching. BM25 (Best Match 25) is the gold-standard algorithm, improving on TF-IDF by incorporating document length normalization and term frequency saturation. Sparse retrieval excels at: exact phrase matching, rare keyword queries, and domain-specific jargon that dense embeddings may not capture well. Implementations typically use inverted indexes for sub-millisecond lookup.

The semantic half of hybrid search. Documents and queries are encoded into high-dimensional embedding vectors (e.g., 768 or 1536 dimensions) using models like text-embedding-3-large or Cohere Embed v3. Similarity is computed via cosine similarity or dot product over an Approximate Nearest Neighbor (ANN) index such as HNSW. Dense retrieval captures paraphrases, conceptual overlap, and cross-lingual semantics that lexical search misses.

The quality of dense retrieval is entirely dependent on the embedding model. Leading options include: OpenAI text-embedding-3-large (3072 dims, MTEB score ~64.6), Cohere Embed v3 (1024 dims, strong multilingual performance), and open-source alternatives like BGE-M3 and E5-mistral-7b-instruct. Critical considerations: maximum token length (often 512-8192 tokens), embedding dimension (higher = more fidelity, more storage), and task-specific fine-tuning on your domain's query-document pairs.

A computationally intensive second-pass stage applied after hybrid retrieval. A cross-encoder model (e.g., Cohere Rerank, BGE-Reranker-v2) processes the query and each candidate document jointly, producing a high-precision relevance score. This corrects the coarse ranking from the fusion stage. Typical pipeline: hybrid retrieval returns top-100 candidates → re-ranker scores and re-orders them → top-10 passed to the LLM context window. Adds ~50-200ms latency but dramatically improves precision@k.

The storage and indexing layer that makes dense retrieval scalable. Leading systems: Pinecone (fully managed, serverless), Weaviate (native hybrid search with BM25 + vector), Qdrant (Rust-based, high-performance HNSW), and Milvus (distributed, cloud-native). Key features for hybrid search: native multi-vector support, built-in RRF fusion, metadata filtering (combining vector search with structured WHERE clauses), and quantization (scalar, product) to reduce memory footprint for billion-scale indexes.