Hybrid Retrieval

Hybrid retrieval combines dense and sparse retrieval, typically by union-then-rerank or reciprocal rank fusion (RRF). The motivation: dense captures semantic similarity, sparse catches exact-token matches; together they cover failure modes neither has alone.

In practice, hybrid often wins by 5–15% NDCG@10 over the best of the two on diverse corpora. The cost is operational — you maintain two indexes and need a fusion strategy.

Most production RAG systems (LlamaIndex, LangChain, Weaviate) ship hybrid as a default option.

A team running a local support-ticket search tool notices two distinct failure patterns: BM25 alone misses paraphrased questions ("my card got declined" vs. a ticket titled "payment authorization failure"), while dense-only search misses exact order IDs and SKU codes. Switching to hybrid retrieval — running both BM25 and a BGE dense index in parallel, then merging with reciprocal rank fusion — fixes both classes of miss without picking a winner in advance. The operational cost is real: two indexes to keep in sync on ingestion, and an extra fusion step before results reach the reranker. For a corpus under a few hundred thousand documents this runs comfortably on CPU alongside a local LLM, since neither BM25 nor HNSW search is GPU-bound.