hybrid search
Also called: hybrid retrieval, hybrid search retrieval, dense-sparse retrieval
Hybrid search is a retrieval method that runs keyword search (BM25) and vector (semantic) search in parallel over the same index, then merges the two ranked lists into one result set, usually with Reciprocal Rank Fusion (RRF). It captures both exact term matches and conceptual meaning.
Keyword search (a sparse method like BM25) tracks individual terms and their frequencies, so it nails exact matches: product codes, jargon, names, dates. Vector search encodes text as dense embeddings and finds passages that are conceptually close even with no shared words. Each fails where the other shines, which is why modern search stacks run both at once.
How the two halves combine
The merge is the tricky part. BM25 scores have no upper limit, while cosine similarity sits between roughly 0.33 and 1.0. You cannot add or average numbers on different scales, so most engines discard the raw scores and fuse by rank instead. Reciprocal Rank Fusion assigns each document 1 / (rank + k) in each list (Azure AI Search uses k = 60), sums those values across lists, and re-sorts. A document that several retrievers rank decently will beat one that a single retriever loves and the others bury.
Some engines expose a weight to tilt the blend, such as Weaviate’s alpha parameter (0 = pure keyword, 1 = pure vector, default 0.75). Hybrid search is now the default retrieval layer inside most RAG and AI-search systems (Azure AI Search, Weaviate, Elastic, OpenSearch) because vendor benchmark testing reports the blend beats either method alone for grounding a generated answer. Understanding it matters because your pages get pulled into an AI answer through exactly this two-sided pipeline, not a single ranking function.
How it affects your traffic
When an AI assistant or an on-site search box answers a query, hybrid retrieval is usually what pulls your page into the candidate set that gets cited or shown. A page that only matches on meaning but omits the literal terms a user types (or the reverse) can rank fine on one half and still get buried after fusion. Our AI SEO work structures your pages to earn a strong rank on both halves at once: exact entity names, product terms and phrasings for the BM25 side, and clear self-contained passages for the vector side, so you surface in the fused results that LLMs draw from.
Get AI SEO that moves the needle
We turn terms like this into ranked pages and qualified pipeline. Start with a free Initial SEO Strategy.