Implement information extraction solutions 5.1 Build retrieval and grounding pipelines

You enable semantic ranker on hybrid queries in Azure AI Search.

Study guide skill: Configure semantic search, hybrid search, and vector search for grounding

Question B1-045

You enable semantic ranker on hybrid queries in Azure AI Search. Which two capabilities does it add? (Choose two.)

Select 2 answers.

  • A Automatic chunking of documents during indexing
  • B Secondary (L2) ranking of the initial BM25- or RRF-ranked results by using language understanding models ✓ Correct answer
  • C Generation of embeddings for every document at indexing time
  • D Semantic captions, and optionally answers, extracted from the top results ✓ Correct answer
  • E Translation of queries into English

Answer: B, D

Semantic ranker reranks an initial BM25- or RRF-scored result set using multilingual deep learning models, and it returns captions and optionally answers. Embeddings and chunking come from skills or integrated vectorization at indexing time, not from semantic ranker.

Deep dive

How it works

Semantic ranker is a set of query-side capabilities that improve an initial result set ranked by BM25 or RRF. It works with text queries, the text portion of vector queries, and hybrid queries. It extends the query pipeline in three ways:

  1. L2 (secondary) ranking. It always adds secondary ranking over the initial results, using multilingual deep learning models adapted from Microsoft Bing to promote the most semantically relevant results.
  2. Captions and answers. It returns captions, which are verbatim sentences and phrases with highlights that best summarize a result, and it can optionally extract answers when a query looks like a question.
  3. Query rewrite (preview). If you turn it on, it creates up to 10 variants of the query, for example fixing typos or rephrasing with generated synonyms.

To use it, you add a semantic configuration to the index. Semantic ranking applies to string (nonvector) fields and works best on informational or descriptive content. Semantic ranker is a premium feature billed by usage, but you can use it for free within the limits of the free tier. With hybrid queries, set k to 50 so semantic ranker gets the most inputs.

The other options, and when you'd use them

  • Automatic chunking during indexing: Chunking comes from the indexing pipeline, through a Text Split skill, the Azure Content Understanding skill or a document parsing mode in the skillset, as part of integrated vectorization.
  • Embedding generation at indexing time: Embeddings come from an embedding skill in the skillset, such as the AzureOpenAIEmbedding skill. Semantic ranker is a set of query-side capabilities that reranks results.
  • Translating queries into English: Semantic ranker's models are multilingual, and query rewrite (preview) creates rephrased variants of a query. For multilingual content in an embedding space, vector queries can find matches without language analyzers or translation.

Exam tip

Semantic ranker adds reranking, captions, answers and (preview) query rewrite on top of BM25 or RRF results. Anything that happens at indexing time, such as chunking or embeddings, belongs to the skillset.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Information extraction questions · AI-103 study guide

Thanks for your feedback!