Add semantic ranking to queries in Azure AI Search

Article
09/25/2024

You can apply semantic ranking to text queries, hybrid queries, and vector queries if your search documents contain string fields and the vector query has a text representation in the search document.

This article explains how to invoke the semantic ranker on queries.

Prerequisites

A search service, Basic tier or higher, with semantic ranker enabled.
An existing search index with a semantic configuration and rich text content.
Review semantic ranking if you need an introduction to the feature.

Note

Captions and answers are extracted verbatim from text in the search document. The semantic subsystem uses machine reading comprehension to recognize content having the characteristics of a caption or answer, but doesn't compose new sentences or phrases. For this reason, content that includes explanations or definitions work best for semantic ranking. If you want chat-style interaction with generated responses, see Retrieval Augmented Generation (RAG).

Choose a client

You can use any of the following tools and SDKs to build a query that uses semantic ranking:

Azure portal, using the index designer to add a semantic configuration.
Visual Studio Code with a REST client
Azure SDK for .NET
Azure SDK for Python
Azure SDK for Java
Azure SDK for JavaScript

Avoid features that bypass relevance scoring

A few query capabilities bypass relevance scoring, which makes them incompatible with semantic ranking. If your query logic includes the following features, you can't semantically rank your results:

A query with search=* or an empty search string, such as pure filter-only query, won't work because there's nothing to measure semantic relevance against and so the search scores are zero. The query must provide terms or phrases that can be evaluated during processing.
Sorting (orderBy clauses) on specific fields overrides search scores and a semantic score. Given that the semantic score is supposed to provide the ranking, adding an orderby clause results in an HTTP 400 error if you apply semantic ranking over ordered results.

Set up the query

In this step, add parameters to the query request. To be successful, your query should be full text search (using the search parameter to pass in a string) or a vector query, and the index should contain text fields with rich semantic content and a semantic configuration.

Search explorer includes options for semantic ranking.

Sign in to the Azure portal.
Open a search index and select Search explorer.
Select Query options. If you already defined a semantic configuration, it's selected by default. If you don't have one, create a semantic configuration for your index.
Enter a query, such as "historic hotel with good food", and select Search.

Alternatively, select JSON view and paste definitions into the query editor:

Here's a JSON example that you can paste into the view:

{
  "search": "funky or interesting hotel with good food on site",
  "count": true,
  "queryType": "semantic",
  "semanticConfiguration": "my-semantic-config",
  "captions": "extractive|highlight-true",
  "answers": "extractive|count-3",
  "queryLanguage": "en-us",
  "highlightPreTag": "<strong>",
  "highlightPostTag": "</strong>",
  "select": "HotelId,HotelName,Description,Category"
}

Use Search Documents to formulate the request.

A response includes an @search.rerankerScore automatically. If you want captions or answers in the response, add captions and answers to the request.

The following example in this section uses the hotels-sample-index to demonstrate semantic ranking with semantic answers and captions.

Paste the following request into a web client as a template. Replace search-service-name with your search service name and replace hotels-sample-index if you have a different index name.

POST https://[search-service-name].search.windows.net/indexes/hotels-sample-index/docs/search?api-version=2024-07-01      
{
      "search": "interesting hotel with restaurant on site and cozy lobby or shared area",
      "count": true,
      "queryType": "semantic",
      "semanticConfiguration": "semantic-config",
      "captions": "extractive|highlight-true",
      "answers": "extractive|count-3",
      "queryLanguage": "en-us",
      "highlightPreTag": "<strong>",
      "highlightPostTag": "</strong>",
      "select": "HotelId,HotelName,Description,Category"
}

Set "queryType" to "semantic".
Set "search" to a full text search query. Your search string can support either the simple syntax or full Lucene syntax. Semantic ranking is an extension of full text search, so while "search" isn't required, you won't get an expected outcome if it's an empty search ("search": "*").
Set "semanticConfiguration" to a predefined semantic configuration that's embedded in your index.
Set "answers" to specify whether semantic answers are included in the result. Currently, the only valid value for this parameter is extractive. Answers can be configured to return a maximum of 10. The default is one. This example shows a count of three answers: extractive|count-3.

Answers aren't guaranteed on every request. To get an answer, the query must look like a question and the content must include text that looks like an answer.
Set "captions" to specify whether semantic captions are included in the result. Currently, the only valid value for this parameter is extractive. Captions can be configured to return results with or without highlights. The default is for highlights to be returned. This example returns captions without highlights: extractive|highlight-false.

The basis for captions and answers are the fields referenced in the "semanticConfiguration". These fields are under a combined limit in the range of 2,000 tokens or approximately 20,000 characters. If you anticipate a token count exceeding this limit, consider a data chunking step using the Text split skill. This approach introduces a dependency on an AI enrichment pipeline and indexers.
Set "highlightPreTag" and "highlightPostTag" if you want to override the default highlight formatting that's applied to captions.

Captions apply highlight formatting over key passages in the document that summarize the response. The default is <em>. If you want to specify the type of formatting (for example, yellow background), you can set the highlightPreTag and highlightPostTag.
Set "select" to specify which fields are returned in the response, and "count" to return the number of matches in the index. These parameters improve the quality of the request and readability of the response.
Send the request to execute the query and return results.

Use QueryType or SemanticQuery to invoke semantic ranker on a semantic query. The following example is from the Azure SDK team.

SearchResults<Hotel> response = await searchClient.SearchAsync<Hotel>(
    "interesting hotel with restaurant on site and cozy lobby or shared area",
    new SearchOptions
    {
        SemanticSearch = new()
        {
            SemanticConfigurationName = "my-semantic-config",
            QueryCaption = new(QueryCaptionType.Extractive),
            QueryAnswer = new(QueryAnswerType.Extractive)
        },
        QueryLanguage = QueryLanguage.EnUs,
        QueryType = SearchQueryType.Semantic
    });

int count = 0;
Console.WriteLine($"Semantic Search Results:");

Console.WriteLine($"\nQuery Answer:");
foreach (QueryAnswerResult result in response.SemanticSearch.Answers)
{
    Console.WriteLine($"Answer Highlights: {result.Highlights}");
    Console.WriteLine($"Answer Text: {result.Text}");
}

await foreach (SearchResult<Hotel> result in response.GetResultsAsync())
{
    count++;
    Hotel doc = result.Document;
    Console.WriteLine($"{doc.HotelId}: {doc.HotelName}");

    if (result.SemanticSearch.Captions != null)
    {
        var caption = result.SemanticSearch.Captions.FirstOrDefault();
        if (caption.Highlights != null && caption.Highlights != "")
        {
            Console.WriteLine($"Caption Highlights: {caption.Highlights}");
        }
        else
        {
            Console.WriteLine($"Caption Text: {caption.Text}");
        }
    }
}
Console.WriteLine($"Total number of search results:{count}");

Evaluate the response

Only the top 50 matches from the initial results can be semantically ranked. As with all queries, a response is composed of all fields marked as retrievable, or just those fields listed in the select parameter. A response includes the original relevance score, and might also include a count, or batched results, depending on how you formulated the request.

In semantic ranking, the response has more elements: a new semantically ranked relevance score, an optional caption in plain text and with highlights, and an optional answer. If your results don't include these extra elements, then your query might be misconfigured. As a first step towards troubleshooting the problem, check the semantic configuration to ensure it's specified in both the index definition and query.

In a client app, you can structure the search page to include a caption as the description of the match, rather than the entire contents of a specific field. This approach is useful when individual fields are too dense for the search results page.

The response for the above example query ("interesting hotel with restaurant on site and cozy lobby or shared area") returns three answers ("answers": "extractive|count-e"). Captions are returned because the "captions" property is set, with plain text and highlighted versions. If an answer can't be determined, it's omitted from the response. For brevity, this example shows just the three answers and the three highest scoring results from the query.

{
  "@odata.count": 29,
  "@search.answers": [
    {
      "key": "24",
      "text": "Chic hotel near the city. High-rise hotel in downtown, within walking distance to theaters, art galleries, restaurants and shops. Visit Seattle Art Museum by day, and then head over to Benaroya Hall to catch the evening's concert performance.",
      "highlights": "Chic hotel near the city. <strong>High-rise hotel in downtown, </strong>within<strong> walking distance to </strong>theaters, art<strong> galleries, restaurants and shops.</strong> Visit Seattle Art Museum by day, and then head over to Benaroya Hall to catch the evening's concert performance.",
      "score": 0.9340000152587891
    },
    {
      "key": "40",
      "text": "Only 8 miles from Downtown. On-site bar/restaurant, Free hot breakfast buffet, Free wireless internet, All non-smoking hotel. Only 15 miles from airport.",
      "highlights": "Only 8 miles from Downtown. <strong>On-site bar/restaurant, Free hot breakfast buffet, Free wireless internet, </strong>All non-smoking<strong> hotel.</strong> Only 15 miles from airport.",
      "score": 0.9210000038146973
    },
    {
      "key": "38",
      "text": "Nature is Home on the beach. Explore the shore by day, and then come home to our shared living space to relax around a stone fireplace, sip something warm, and explore the library by night. Save up to 30 percent. Valid Now through the end of the year. Restrictions and blackouts may apply.",
      "highlights": "Nature is Home on the beach. Explore the shore by day, and then come home to our<strong> shared living space </strong>to relax around a stone fireplace, sip something warm, and explore the library by night. Save up to 30 percent. Valid Now through the end of the year. Restrictions and blackouts may apply.",
      "score": 0.9200000166893005
    }
  ],
  "value": [
    {
      "@search.score": 3.2328331,
      "@search.rerankerScore": 2.575303316116333,
      "@search.captions": [
        {
          "text": "The best of old town hospitality combined with views of the river and cool breezes off the prairie. Our penthouse suites offer views for miles and the rooftop plaza is open to all guests from sunset to 10 p.m. Enjoy a complimentary continental breakfast in the lobby, and free Wi-Fi throughout the hotel.",
          "highlights": "The best of old town hospitality combined with views of the river and cool breezes off the prairie. Our<strong> penthouse </strong>suites offer views for miles and the rooftop<strong> plaza </strong>is open to all guests from sunset to 10 p.m. Enjoy a<strong> complimentary continental breakfast in the lobby, </strong>and free Wi-Fi<strong> throughout </strong>the hotel."
        }
      ],
      "HotelId": "50",
      "HotelName": "Head Wind Resort",
      "Description": "The best of old town hospitality combined with views of the river and cool breezes off the prairie. Our penthouse suites offer views for miles and the rooftop plaza is open to all guests from sunset to 10 p.m. Enjoy a complimentary continental breakfast in the lobby, and free Wi-Fi throughout the hotel.",
      "Category": "Suite"
    },
    {
      "@search.score": 0.632956,
      "@search.rerankerScore": 2.5425150394439697,
      "@search.captions": [
        {
          "text": "Every stay starts with a warm cookie. Amenities like the Counting Sheep sleep experience, our Wake-up glorious breakfast buffet and spacious workout facilities await.",
          "highlights": "Every stay starts with a warm cookie. Amenities like the<strong> Counting Sheep sleep experience, </strong>our<strong> Wake-up glorious breakfast buffet and spacious workout facilities </strong>await."
        }
      ],
      "HotelId": "34",
      "HotelName": "Lakefront Captain Inn",
      "Description": "Every stay starts with a warm cookie. Amenities like the Counting Sheep sleep experience, our Wake-up glorious breakfast buffet and spacious workout facilities await.",
      "Category": "Budget"
    },
    {
      "@search.score": 3.7076726,
      "@search.rerankerScore": 2.4554927349090576,
      "@search.captions": [
        {
          "text": "Chic hotel near the city. High-rise hotel in downtown, within walking distance to theaters, art galleries, restaurants and shops. Visit Seattle Art Museum by day, and then head over to Benaroya Hall to catch the evening's concert performance.",
          "highlights": "Chic hotel near the city. <strong>High-rise hotel in downtown, </strong>within<strong> walking distance to </strong>theaters, art<strong> galleries, restaurants and shops.</strong> Visit Seattle Art Museum by day, and then head over to Benaroya Hall to catch the evening's concert performance."
        }
      ],
      "HotelId": "24",
      "HotelName": "Uptown Chic Hotel",
      "Description": "Chic hotel near the city. High-rise hotel in downtown, within walking distance to theaters, art galleries, restaurants and shops. Visit Seattle Art Museum by day, and then head over to Benaroya Hall to catch the evening's concert performance.",
      "Category": "Suite"
    },
   . . .
  ]
}

Expected workloads

For semantic ranking, you should expect a search service to support up to 10 concurrent queries per replica.

The service throttles semantic ranking requests if volumes are too high. An error message that includes these phrases indicate the service is at capacity for semantic ranking:

Error in search query: Operation returned an invalid status 'Partial Content'`
@search.semanticPartialResponseReason`
CapacityOverloaded

If you anticipate consistent throughput requirements near, at, or higher than this level, please file a support ticket so that we can provision for your workload.

Next steps

Semantic ranking can be used in hybrid queries that combine keyword search and vector search into a single request and a unified response.

Hybrid query with semantic ranker

Share via