Hybrid Search

The keyword-based backbone of hybrid search, using the BM25 algorithm to match exact query terms against an inverted index.

• Excels at matching rare, specific terms like product codes, technical jargon, and proper nouns
• Uses TF-IDF weighting to prioritize documents where query terms appear frequently but are rare across the corpus
• Handles out-of-domain queries reliably without requiring training data
• Fails on semantic gaps: a search for 'cardiac arrest' won't match 'heart attack' without synonym expansion

BM25 remains the gold standard for precision when exact term matching is critical, such as in legal document retrieval or medical literature search.

The semantic understanding layer that encodes queries and documents into high-dimensional embedding vectors using transformer models.

• Captures conceptual relationships: 'inexpensive laptop' retrieves documents about 'budget notebooks' even without keyword overlap
• Uses cosine similarity or dot product to measure vector proximity in embedding space
• Requires an embedding model (e.g., text-embedding-3-large, Cohere Embed v3) and a vector database for efficient ANN search
• Vulnerable to domain shift: embeddings trained on general text may underperform on specialized enterprise jargon

Dense retrieval shines when users express intent through natural language rather than precise keywords, making it essential for conversational AI interfaces.

The score normalization algorithm that merges sparse and dense result sets into a single ranked list without requiring model calibration.

• Formula: RRFscore(d) = Σ 1/(k + rank_i(d)) where k is a constant (typically 60)
• Rank-agnostic: works with any retrieval system regardless of internal scoring mechanics
• Outperforms linear combination when score distributions differ dramatically between retrievers
• Computationally lightweight compared to learning-to-rank fusion models

RRF is the industry-standard fusion method because it requires no training data and produces consistently strong results across diverse retrieval scenarios.

A second-pass refinement stage that applies a transformer model to jointly encode the query and each candidate document for fine-grained relevance scoring.

• Unlike bi-encoders (which encode queries and documents separately), cross-encoders perform full attention between query-document pairs
• Dramatically improves precision at the cost of higher latency: reranking top-100 documents is feasible, reranking millions is not
• Models like Cohere Rerank and BGE-Reranker are purpose-built for this task
• Essential for applications where precision@10 is the critical metric, such as RAG systems feeding context into LLMs

Cross-encoder reranking is the quality gate that ensures only the most relevant documents survive into the final context window.

The document segmentation methodology that determines retrieval granularity and directly impacts hybrid search effectiveness.

• Fixed-size chunks (e.g., 512 tokens with 10% overlap) provide simplicity but may split coherent concepts
• Semantic chunking uses sentence boundaries and embedding similarity to create conceptually coherent segments
• Hierarchical chunking maintains parent-child relationships: retrieve small chunks for precision, return larger parent chunks for context
• Chunk size represents a fundamental tradeoff: smaller chunks improve retrieval precision, larger chunks preserve context for downstream generation

Effective chunking is the unsung hero of hybrid search performance, determining whether the right information can be found and surfaced.

The structured constraint layer that applies boolean filters on document metadata before or alongside vector similarity search.

• Pre-filtering: apply metadata constraints first, then perform vector search on the reduced set — guarantees filter compliance but may miss relevant results if the filter is too aggressive
• Post-filtering: perform vector search first, then remove non-matching documents — preserves semantic recall but may return fewer results than requested
• Custom hybrid filtering in vector databases like Weaviate and Pinecone allows filter-aware ANN indices
• Critical for enterprise use cases: 'find documents about revenue projections from Q3 2024 authored by the finance team'

Metadata filtering bridges the gap between unstructured semantic search and structured database queries.

A bag-of-words retrieval function that ranks documents by term frequency saturation and inverse document frequency. BM25 excels at exact keyword matching and handles rare terms well, but fails on synonyms and paraphrases.

• Based on the probabilistic relevance framework
• Uses non-linear term frequency saturation to prevent keyword stuffing
• Default sparse retrieval component in most hybrid search pipelines
• Computationally lightweight compared to dense embeddings

A retrieval method where queries and documents are encoded into fixed-length embedding vectors using transformer models. Relevance is computed via cosine similarity or dot product in high-dimensional space.

• Captures semantic meaning beyond exact keywords
• Requires an approximate nearest neighbor (ANN) index for scalability
• Embedding models include text-embedding-3, bge-large, and E5
• Sensitive to domain shift between training data and target corpus

A score normalization and merging algorithm that combines ranked lists from multiple retrieval sources without requiring calibrated scores. RRF assigns a score of 1 / (k + rank) to each document and sums across sources.

• Parameter k controls the decay rate (typically k=60)
• Outperforms min-max normalization when score distributions differ
• Used in Elasticsearch 8.8+ and OpenSearch 2.12+
• Stateless and deterministic — no training required

The system design pattern where BM25 and vector search run in parallel, and results are merged via RRF or a learned combiner. This architecture ensures both lexical precision and semantic recall.

• Two-stage retrieval: candidate generation (sparse + dense) → reranking
• Can include metadata filtering (date, author, category) as a third constraint
• Common implementations: Weaviate, Pinecone, Vespa, Elasticsearch
• Trade-off: doubled index storage and query latency vs. single-method retrieval

A middle-ground approach where a neural model generates contextualized term weights rather than dense vectors. Methods like SPLADE and uniCOIL produce sparse vectors that behave like learned BM25 weights.

• Retains inverted index compatibility for fast retrieval
• Captures term expansion (synonyms, related terms) automatically
• Outperforms BM25 on semantic benchmarks while maintaining interpretability
• Can serve as the sparse leg in hybrid systems without a separate dense index

A second-stage refinement where a transformer model processes the query and each candidate document together as a pair, outputting a precise relevance score. Unlike bi-encoders, cross-encoders attend to query-document interactions.

• Models: BGE-reranker, Cohere Rerank, MonoT5
• Applied to top-k candidates (typically 100-200) from first-stage hybrid retrieval
• Significantly improves precision@k at the cost of per-pair inference latency
• Essential for production RAG systems requiring high answer fidelity