Hybrid Search for RAG: BM25, SPLADE, and Vector Search Combined

Hybrid search is a method that combines both sparse and dense retrieval techniques to enhance search precision and recall, especially in scenarios where vocabulary mismatch occurs between queries and documents. It involves using a sparse retriever like BM25 or SPLADE alongside a dense vector retriever, merging the results with fusion algorithms such as Reciprocal Rank Fusion (RRF) or Convex Combination, and optionally employing a cross-encoder reranker for additional precision. The approach is particularly beneficial in domain-specific Retrieval-Augmented Generation (RAG) systems where queries include both exact identifiers and semantic intent. While hybrid search consistently outperforms single-method retrievals by covering the weaknesses of both dense and sparse methods, it requires careful tuning and evaluation on real user data to ensure improvement over dense-only setups. Vector databases like Qdrant, Weaviate, and Elasticsearch have varying levels of support for hybrid search, offering different fusion methods and customization options to suit diverse retrieval scenarios.