Hybrid Search
Hybrid search is a technique that combines keyword-based full-text search (such as BM25) with vector search (semantic search), leveraging the strengths of both to improve retrieval accuracy.
The single biggest factor in RAG retrieval accuracy is whether relevant documents are correctly retrieved. Vector search alone struggles with exact matches for proper nouns like "ISO 27001," while BM25 alone cannot handle semantic paraphrases like "international standard for information security." Hybrid search compensates for these two weaknesses. The most common implementation pattern runs BM25 and vector search independently, then merges results using RRF (Reciprocal Rank Fusion). It sums the reciprocals of each method's rankings to produce a final reranked score. The formula is simple, yet it consistently outperforms either search method alone. Combining pgvector with PostgreSQL's full-text search enables hybrid search without additional infrastructure. On Supabase, maintaining both a vector column and a tsvector column in the same table and computing both scores within SQL is a practical approach. Chunk size design also affects accuracy. Smaller chunks improve vector search precision but lose context for BM25. In practice, chunks of 500-1000 tokens with overlapping context from adjacent segments are common.
Related Terms
BM25
BM25 (Best Matching 25) is a probabilistic information retrieval algorithm that scores the relevance between a query and documents by taking into account the term frequency within documents and document length.
RRF
RRF (Reciprocal Rank Fusion) is a scoring method that integrates ranking results returned by multiple retrieval methods. By summing the reciprocal ranks from each method, it enables the fusion of different scoring systems without normalization.
RAG (Retrieval-Augmented Generation)
RAG (Retrieval-Augmented Generation) is a technique that improves the accuracy and currency of responses by retrieving relevant information from external knowledge sources and appending the results to the input of an LLM.
