In short. Hybrid retrieval combines term matching with vector similarity and merges their results into a ranked list.

Definition: Hybrid retrieval combines lexical retrieval, which matches query terms against indexed text, with vector retrieval, which searches numerical representations called embeddings for similar content. It merges their results into a ranked list.

Key facts

As of 17 October 2026, Microsoft documents the following behavior in Azure AI Search:

Microsoft's hybrid search overview describes a single request that runs full-text and vector queries in parallel.
Microsoft's relevance scoring documentation describes reciprocal rank fusion (RRF), which combines results using their positions in separate ranked lists.
Microsoft's semantic ranking overview describes a subsequent text-based reranking step that processes only the top 50 results.

Why combine lexical and vector retrieval?

Microsoft's hybrid search overview identifies product codes and people's names as cases where keyword search can identify exact matches. Vector search can find conceptually similar information even when the query's words are absent from the indexed text. Combining the methods retains access to both kinds of match.

For a website's own search application, the vector search overview describes storing text and vector content in the same Azure index. Embeddings can be generated externally and loaded into vector fields, or generated through an integrated indexing pipeline. The query also needs a vector representation for similarity search.

How are the result lists combined?

Microsoft's RRF documentation explains that each query first produces a ranked list. RRF assigns each document a contribution based on its position in each list, adds those contributions, and orders the combined results. A document appearing high in several lists receives contributions from each of them.

Semantic reranking is an optional later operation. Microsoft's semantic ranking overview explains that the reranker reads text from the initial results and computes new relevance scores. Its candidate set consists of the top 50 retrieved results; it cannot retrieve additional documents from the whole collection during this reranking step.

What does a hybrid query look like?

The following request-body excerpt is adapted from Microsoft's hybrid query example. The embedding placeholder requires an actual query embedding before execution against an index with the named vector field.

{
  "search": "historic hotel walk to restaurants and shopping",
  "vectorQueries": [
    {
      "kind": "vector",
      "vector": [ <array of embeddings> ],
      "k": 50,
      "fields": "DescriptionVector"
    }
  ]
}

The search property supplies the lexical query, while vectorQueries supplies the vector query. The example combines retrieval methods without requesting the optional semantic reranker.

FAQ

Does hybrid retrieval require embeddings?

Yes. Microsoft's vector search documentation describes encoding both source content and queries as vectors. In Azure, you can supply embeddings externally or use integrated vectorization.

How can you inspect the contribution of each retrieval method?

Microsoft's scoring documentation describes adding "debug": "vector" to a query request to return subscores. These help inspect vector contributions when assessing a hybrid query's ranking.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer