What is hybrid search?

What is hybrid search and how do the two sides combine?

A retrieval pipeline runs two indexes over the same corpus and asks both at once. The keyword index scores documents with BM25 or a learned sparse encoder such as SPLADE, so it rewards documents that contain the query tokens themselves, including rare terms, product codes, and acronyms the embedding model has never seen. The dense index scores documents by the cosine similarity between a query embedding and a document embedding, so it rewards semantic overlap even when no words match. Each side returns its own top-k. Hybrid search is the fusion step that turns those two ranked lists into one.

Every hosted retriever exposes this as a single call. Weaviate's hybrid() query takes an alpha in [0, 1], where 1 is pure vector and 0 is pure keyword, and a fusion type of ranked or relative score. Elasticsearch exposes an rrf retriever that takes a list of child retrievers, a rank_window_size, and a rank_constant. Pinecone stores dense and sparse vectors on the same record and combines them server-side, or runs the two searches separately and fuses the results client-side with RRF. The shape is the same in every case: two rankers, one fused list.

Reciprocal Rank Fusion vs a weighted score sum

RRF ignores the raw scores and works on ranks. For each document that appears in either list, the fused score is the sum over lists of 1 / (k + rank), where rank starts at 1 and k is a smoothing constant. Cormack, Clarke, and Buettcher's 2009 SIGIR paper set k = 60, and Elasticsearch and most reference implementations kept that default. The appeal of RRF is that BM25 scores and cosine similarities live on incompatible scales, and a rank-only method sidesteps that problem without a normalisation pass over the corpus.

Weighted score sums do combine the raw scores, usually as alpha * dense_score + (1 - alpha) * sparse_score. That requires normalising both sides first, since BM25 is unbounded and cosine sits in [-1, 1]. Weaviate's relative score fusion is exactly this shape: it min-max normalises each side to [0, 1] and blends them with the alpha the caller passed. The benefit is that a strong keyword score dominates the fused list even when its rank sits close to the dense rank; the cost is one more tuning knob and a normalisation step that has to hold up as the corpus grows.

When is hybrid search worth two indexes?

Hybrid pays off when the query distribution mixes exact-token lookups and natural-language questions over the same corpus. A support portal that answers both ERR_CONNECTION_RESET and "why can't my laptop reach the wifi" is the canonical case: BM25 wins the first query, the dense retriever wins the second, and neither wins both on its own. Product catalogues with SKUs and free-text descriptions, code search over identifiers and comments, and legal or medical corpora full of statute numbers and paraphrase all show the same pattern.

The cost is that you now maintain two indexes, run two searches per query, and tune a fusion parameter on top. If your queries are all natural language over a corpus with no jargon, a dense retriever alone is usually within a point or two of the hybrid version and simpler to run. If your queries are all keyword lookups over structured records, BM25 alone stays cheaper and easier to explain. Measure recall at 50 and precision at 5 on a labelled query set from your own traffic before adding a second index.

Hybrid search is orthogonal to reranking. Fusion picks the best top-k across two first-stage retrievers; a reranker then rescores each candidate against the query with a cross-encoder to reorder that top-k. A common production pipeline runs BM25 and dense retrieval, fuses the lists with RRF, and reranks the fused top-k before the LLM sees it.

POST /docs/_search
{
  "retriever": {
    "rrf": {
      "retrievers": [
        { "standard": {
            "query": { "match": { "text": "rotate signing key without downtime" } }
        } },
        { "knn": {
            "field": "text_vector",
            "query_vector": [0.12, -0.03, 0.44, /* ... 764 more dims */],
            "k": 50,
            "num_candidates": 200
        } }
      ],
      "rank_window_size": 50,
      "rank_constant": 60
    }
  },
  "size": 5
}