Hybrid Search - Mixed Search Strategy

Basic Information

• Type: Technical Strategy/Architectural Pattern
• Domain: Information Retrieval, RAG Optimization
• Core Idea: Combine the advantages of lexical retrieval and semantic retrieval
• Fusion Methods: RRF (Reciprocal Rank Fusion), Linear Combination

Concept Description

Hybrid Search is a retrieval strategy that combines lexical retrieval (e.g., BM25) and semantic retrieval (e.g., vector search). It runs both retrievers in parallel, merges the results using a fusion algorithm, and optionally applies re-ranking. This method leverages the strengths of precise keyword matching and semantic understanding, widely regarded as the best retrieval practice in RAG systems.

Core Principles

1. Parallel Retrieval: BM25 and vector search run simultaneously
2. Result Fusion: Merge the two sets of results using RRF or linear weighting
3. Optional Re-ranking: Further refine using Cross-Encoder
4. Pass to LLM: Provide the final results as context to the generative model

Why Hybrid Search is Needed

Performance Improvement

• Recall improves from ~0.72 (BM25) to ~0.91 (Hybrid)
• Precision improves from ~0.68 (BM25) to ~0.87 (Hybrid)
• NVIDIA reports hybrid architecture achieves 96% factual fidelity on financial documents
• In Anthropic Contextual RAG, contextual embeddings + BM25 reduce retrieval failure rate by 49%

Fusion Algorithms

RRF (Reciprocal Rank Fusion)

• Formula: RRF_score = Σ 1/(k + rank_i)
• Advantage: No need for tuning, robust
• Applicability: When the score scales of the two retrievers are inconsistent

Linear Combination

• Formula: score = α × BM25_score + (1-α) × vector_score
• Advantage: Tunable weights, flexible
• Applicability: When fine control over the contributions of both retrievers is needed

Advanced Variants

• Hybrid + Reranker: Hybrid retrieval + Cross-Encoder re-ranking (best practice)
• Hybrid + SPLADE: Three-way fusion of BM25 + SPLADE + Dense
• Graph + Vector: Hybrid of GraphRAG + vector retrieval
• Contextual Hybrid: Anthropic's contextual embeddings + contextual BM25

Applicable Scenarios

• Queries containing exact identifiers (error codes, product names, API endpoints, legal terms) and natural language intent
• Document sets containing technical documents and natural language descriptions
• Need to handle both known terms and vague queries
• High-precision enterprise search

Mainstream Implementations

Relationship with OpenClaw Ecosystem

Hybrid search is the recommended retrieval strategy for the OpenClaw RAG pipeline. By combining BM25 (exact matching of user terms and keywords) and vector search (understanding user intent), OpenClaw can provide more comprehensive and accurate retrieval results. Especially in personal knowledge bases, users may search for specific file names, person names (BM25 excels) or vague thematic concepts (vector search excels), and hybrid search can meet both types of needs.