In short. Semantic search uses query meaning to find or rank content, with vector retrieval and semantic reranking operating at different stages.

Definition: Semantic search retrieves or ranks information according to the meaning of a query and its relationship to content. It can use numerical representations to retrieve similar content and language models to rerank an existing result set.

Key facts

As of 17 October 2026, Microsoft's semantic ranking documentation describes Azure AI Search's semantic ranker as a secondary ranking stage that processes the top 50 retrieved results.
Microsoft's vector search overview describes matching numerical representations of content by similarity, including conceptually related words such as dog and canine.
Microsoft's hybrid search overview combines keyword and vector retrieval in one request, merging their results with Reciprocal Rank Fusion, a method for combining ranked result lists.
Microsoft's query instructions require an existing index with a semantic configuration and rich text content to use semantic ranking.

How does meaning-based retrieval find candidates?

An embedding is a numerical representation of content produced by a model. In Microsoft's vector retrieval workflow, the documents and query are encoded as vectors, and similarity search retrieves documents whose vectors are close to the query vector. The dog and canine example illustrates how related concepts can match despite different wording.

Keyword retrieval remains useful alongside this approach. Microsoft's hybrid search documentation identifies product codes and people's names as cases where exact matching can perform better. Its hybrid workflow runs keyword and vector searches in parallel, then merges their ranked lists. Semantic reranking can follow this merged candidate set.

What does semantic reranking change?

Reranking assigns new relevance scores to results already retrieved. Microsoft's semantic ranker overview describes models that evaluate query context against text from the candidate documents. Its example uses the word capital, whose meaning depends on whether the context concerns finance, law, geography or grammar.

In Azure AI Search, only the top 50 candidates reach this stage. The ranker works on text, including text associated with vector results. It returns an @search.rerankerScore from 0 to 4, with higher values indicating greater relevance to the query. A relevant document outside that candidate set stays outside the reranking step.

For a website's own search implementation, descriptive prose supplies material for this processing. The same documentation identifies knowledge bases and online documentation as suitable content. Candidate retrieval and subsequent ordering therefore remain separate parts of the implementation.

How can semantic ranking be checked?

Microsoft's query setup guide provides a hotel-search example. This reduced request body retains its query and configuration settings:

{
  "search": "funky or interesting hotel with good food on site",
  "queryType": "semantic",
  "semanticConfiguration": "my-semantic-config",
  "captions": "extractive|highlight-true"
}

The configuration name must identify a semantic configuration defined in the index. The response includes @search.rerankerScore; the captions setting requests extracted passages with highlighting. The same guide recommends checking that the configuration appears in both the index definition and the query when expected response elements are missing.

FAQ

Do I need vectors to use semantic ranking?

You can apply Azure's semantic ranking to text queries, according to Microsoft's query guide. Vector and hybrid queries can also use it when the search documents contain the required text representation.

Does a semantic answer contain newly written text?

Azure's semantic answers extract passages verbatim from indexed content, according to Microsoft's semantic ranker overview. An answer requires a question-like query and a passage that supplies a likely answer.

Sources

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