Hybrid search combines BM25 lexical retrieval and dense vector retrieval in parallel, then fuses their ranked results into a single list. BM25 excels at exact identifiers, error codes, and rare terms; dense vectors handle paraphrase, synonymy, and conversational queries. The two main fusion methods are Reciprocal Rank Fusion (RRF), which is rank-based and needs no normalization, and weighted sum, which requires score normalization but exploits score magnitude. OpenSearch implements hybrid via a `hybrid` query type with a normalization-processor pipeline, while Elasticsearch uses the `retriever` API with an `rrf` retriever (GA in 8.16). The post covers weight tuning, evaluation metrics (nDCG@10, Recall@100, MRR), score drift risks when re-embedding, and the additive memory cost of running both an inverted index and an HNSW graph.
