Create a hybrid query in Azure AI Search

Artikkeli
10/07/2024

Hybrid search combines text (keyword) and vector queries in a single search request. All subqueries in the request execute in parallel. The results are merged and reordered by new search scores, using Reciprocal Rank Fusion (RRF) to return a unified result set. In many cases, per benchmark tests, hybrid queries with semantic ranking return the most relevant results.

In this article, learn how to:

Set up a basic request
Formulate hybrid queries with more parameters and filters
Improve relevance using semantic ranking or vector weights
Optimize query behaviors by controlling text and vector inputs

Note

New in 2024-09-01-preview is the ability to target filters to just the vector subqueries in a hybrid request. This gives you more precision over how filters are applied. For more information, see targeting filters to vector subqueries in this article.

Prerequisites

A search index containing searchable vector and nonvector fields. We recommend the Import and vectorize data wizard to create an index quickly. Otherwise, see Create an index and Add vector fields to a search index.
(Optional) If you want the semantic ranker, your search service must be Basic tier or higher, with semantic ranker enabled.
(Optional) If you want built-in text-to-vector conversion of a query string, create and assign a vectorizer to vector fields in the search index.

Choose an API or tool

Search Explorer in the Azure portal (supports both stable and preview API search syntax) has a JSON view that lets you paste in a hybrid request.
2024-07-01 stable version or a recent preview API version if you're using preview features like maxTextRecallSize and countAndFacetMode(preview).

For readability, we use REST examples to explain how the APIs work. You can use a REST client like Visual Studio Code with the REST extension to build hybrid queries. For more information, see Quickstart: Vector search using REST APIs.
Newer stable or beta packages of the Azure SDKs (see change logs for SDK feature support).

Set up a hybrid query in Search Explorer

In Search Explorer, make sure the API version is 2024-07-01 or a newer preview API version.
Under View, select JSON view so that you can paste in a vector query.

Replace the default query template with a hybrid query, such as the "Run a hybrid query" example starting on line 539 in the vector quickstart. For brevity, the vector is truncated in this article.

A hybrid query has a text query specified in search, and a vector query specified under vectorQueries.vector.

The text query and vector query can be equivalent or divergent, but it's common for them to share the same intent.

{
    "count": true,
    "search": "historic hotel walk to restaurants and shopping",
    "select": "HotelId, HotelName, Category, Tags, Description",
    "top": 7,
    "vectorQueries": [
        {
            "vector": [0.01944167, 0.0040178085, -0.007816401 ... <remaining values omitted> ], 
            "k": 7,
            "fields": "DescriptionVector",
            "kind": "vector",
            "exhaustive": true
        }
    ]
}

Select Search.

Tip

Search results are easier to read if you hide the vectors. In Query Options, turn on Hide vector values in search results.

Hybrid query request (REST API)

A hybrid query combines text search and vector search, where the search parameter takes a query string and vectorQueries.vector takes the vector query. The search engine runs full text and vector queries in parallel. The union of all matches is evaluated for relevance using Reciprocal Rank Fusion (RRF) and a single result set is returned in the response.

Results are returned in plain text, including vectors in fields marked as retrievable. Because numeric vectors aren't useful in search results, choose other fields in the index as a proxy for the vector match. For example, if an index has "descriptionVector" and "descriptionText" fields, the query can match on "descriptionVector" but the search result can show "descriptionText". Use the select parameter to specify only human-readable fields in the results.

The following example shows a hybrid query configuration.

POST https://{{search-service-name}}.search.windows.net/indexes/{{index-name}}/docs/search?api-version=2024-07-01
Content-Type: application/json
api-key: {{admin-api-key}}
{
    "vectorQueries": [
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "exhaustive": true,
            "k": 10
        },
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "exhaustive": true,
            "k": 10
        }
    ],
    "search": "historic hotel walk to restaurants and shopping",
    "select": "HotelName, Description, Address/City",
    "top": 10
}

Key points:

The vector query string is specified through the vectorQueries.vector property. The query executes against the "DescriptionVector" field. Set kind to "vector" to indicate the query type. Optionally, set exhaustive to true to query the full contents of the vector field.
Keyword search is specified through search property. It executes in parallel with the vector query.
k determines how many nearest neighbor matches are returned from the vector query and provided to the RRF ranker.
top determines how many matches are returned in the response all-up. In this example, the response includes 10 results, assuming there are at least 10 matches in the merged results.

Hybrid search with filter

This example adds a filter, which is applied to the filterable nonvector fields of the search index.

POST https://{{search-service-name}}.search.windows.net/indexes/{{index-name}}/docs/search?api-version=2024-07-01
Content-Type: application/json
api-key: {{admin-api-key}}
{
    "vectorQueries": [
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "k": 10
        }
    ],
    "search": "historic hotel walk to restaurants and shopping",
    "vectorFilterMode": "postFilter",
    "filter": "ParkingIncluded",
    "top": "10"
}

Key points:

Filters are applied to the content of filterable fields. In this example, the ParkingIncluded field is a boolean and it's marked as filterable in the index schema.
In hybrid queries, filters can be applied before query execution to reduce the query surface, or after query execution to trim results. "preFilter" is the default. To use postFilter, set the filter processing mode as shown in this example.
When you postfilter query results, the number of results might be less than top-n.

Hybrid search with filters targeting vector subqueries (preview)

Using 2024-09-01-preview, you can override a global filter on the search request by applying a secondary filter that targets just the vector subqueries in a hybrid request.

This feature provides fine-grained control by ensuring that filters only influence the vector search results, leaving keyword-based search results unaffected.

The targeted filter fully overrides the global filter, including any filters used for security trimming or geospatial search. In cases where global filters are required, such as security trimming, you must explicitly include these filters in both the top-level filter and in each vector-level filter to ensure security and other constraints are consistently enforced.

To apply targeted vector filters:

Use the latest preview Search Documents REST API or an Azure SDK beta package that provides the feature.
Modify a query request, adding a new vectorQueries.filterOverride parameter set to an OData filter expression.

Here's an example of hybrid query that adds a filter override. The global filter "Rating gt 3" is replaced at run time by the filterOvrride.

POST https://{{search-service-name}}.search.windows.net/indexes/{{index-name}}/docs/search?api-version=2024-09-01=preview

{
    "vectorQueries": [
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "exhaustive": true,
            "filterOverride": "Address/City eq 'Seattle'",
            "k": 10
        }
    ],
    "search": "historic hotel walk to restaurants and shopping",
    "select": "HotelName, Description, Address/City, Rating",
    "filter": "Rating gt 3"
    "debug": "vector",
    "top": 10
}

Semantic hybrid search

Assuming that you enabled semantic ranker and your index definition includes a semantic configuration, you can formulate a query that includes vector search and keyword search, with semantic ranking over the merged result set. Optionally, you can add captions and answers.

Whenever you use semantic ranking with vectors, make sure k is set to 50. Semantic ranker uses up to 50 matches as input. Specifying less than 50 deprives the semantic ranking models of necessary inputs.

POST https://{{search-service-name}}.search.windows.net/indexes/{{index-name}}/docs/search?api-version=2024-07-01
Content-Type: application/json
api-key: {{admin-api-key}}
{
    "vectorQueries": [
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "k": 50
        }
    ],
    "search": "historic hotel walk to restaurants and shopping",
    "select": "HotelName, Description, Tags",
    "queryType": "semantic",
    "semanticConfiguration": "my-semantic-config",
    "captions": "extractive",
    "answers": "extractive",
    "top": "50"
}

Key points:

Semantic ranker accepts up to 50 results from the merged response.
"queryType" and "semanticConfiguration" are required.
"captions" and "answers" are optional. Values are extracted from verbatim text in the results. An answer is only returned if the results include content having the characteristics of an answer to the query.

Semantic hybrid search with filter

Here's the last query in the collection. It's the same semantic hybrid query as the previous example, but with a filter.

POST https://{{search-service-name}}.search.windows.net/indexes/{{index-name}}/docs/search?api-version=2024-07-01
Content-Type: application/json
api-key: {{admin-api-key}}
{
    "vectorQueries": [
        {
            "vector": [
                -0.009154141,
                0.018708462,
                . . . 
                -0.02178128,
                -0.00086512347
            ],
            "fields": "DescriptionVector",
            "kind": "vector",
            "k": 50
        }
    ],
    "search": "historic hotel walk to restaurants and shopping",
    "select": "HotelName, Description, Tags",
    "queryType": "semantic",
    "semanticConfiguration": "my-semantic-config",
    "captions": "extractive",
    "answers": "extractive",
    "filter": "ParkingIsIncluded'",
    "vectorFilterMode": "postFilter",
    "top": "50"
}

Key points:

The filter mode can affect the number of results available to the semantic reranker. As a best practice, it's smart to give the semantic ranker the maximum number of documents (50). If prefilters or postfilters are too selective, you might be underserving the semantic ranker by giving it fewer than 50 documents to work with.
Prefiltering is applied before query execution. If prefilter reduces the search area to 100 documents, the vector query executes over the "DescriptionVector" field for those 100 documents, returning the k=50 best matches. Those 50 matching documents then pass to RRF for merged results, and then to semantic ranker.
Postfilter is applied after query execution. If k=50 returns 50 matches on the vector query side, followed by a post-filter applied to the 50 matches, your results are reduced by the number of documents that meet filter criteria. This leaves you with fewer than 50 documents to pass to semantic ranker. Keep this in mind if you're using semantic ranking. The semantic ranker works best if it has 50 documents as input.

Set maxTextRecallSize and countAndFacetMode (preview)

This section explains how to adjust the inputs to a hybrid query by controlling the amount BM25-ranked results that flow to the hybrid ranking model. Controlling over the BM25-ranked input gives you more options for relevance tuning in hybrid scenarios.

We recommend preview REST API version 2024-05-01-preview.

Tip

Another option to consider is a supplemental or replacement technique, is vector weighting, which increases the importance of vector queries in the request.

Use Search - POST or Search - GET in 2024-05-01-preview to specify these parameters.
Add a hybridSearch query parameter object to set the maximum number of documents recalled through the BM25-ranked results of a hybrid query. It has two properties:
- maxTextRecallSize specifies the number of BM25-ranked results to provide to the Reciprocal Rank Fusion (RRF) ranker used in hybrid queries. The default is 1,000. The maximum is 10,000.
- countAndFacetMode reports the counts for the BM25-ranked results (and for facets if you're using them). The default is all documents that match the query. Optionally, you can scope "count" to the maxTextRecallSize.
Reduce maxTextRecallSize if vector similarity search is generally outperforming the text-side of the hybrid query.
Raise maxTextRecallSize if you have a large index, and the default isn't capturing a sufficient number of results. With a larger BM25-ranked result set, you can also set top, skip, and next to retrieve portions of those results.

The following REST examples show two use-cases for setting maxTextRecallSize.

The first example reduces maxTextRecallSize to 100, limiting the text side of the hybrid query to just 100 document. It also sets countAndFacetMode to include only those results from maxTextRecallSize.

POST https://[service-name].search.windows.net/indexes/[index-name]/docs/search?api-version=2024-05-01-Preview 

    { 
      "vectorQueries": [ 
        { 
          "kind": "vector", 
          "vector": [1.0, 2.0, 3.0], 
          "fields": "my_vector_field", 
          "k": 10 
        } 
      ], 
      "search": "hello world", 
      "hybridSearch": { 
        "maxTextRecallSize": 100, 
        "countAndFacetMode": "countRetrievableResults" 
      } 
    }

The second example raises maxTextRecallSize to 5,000. It also uses top, skip, and next to pull results from large result sets. In this case, the request pulls in BM25-ranked results starting at position 1,500 through 2,000 as the text query contribution to the RRF composite result set.

POST https://[service-name].search.windows.net/indexes/[index-name]/docs/search?api-version=2024-05-01-Preview 

    { 
      "vectorQueries": [ 
        { 
          "kind": "vector", 
          "vector": [1.0, 2.0, 3.0], 
          "fields": "my_vector_field", 
          "k": 10 
        } 
      ], 
      "search": "hello world",
      "top": 500,
      "skip": 1500,
      "next": 500,
      "hybridSearch": { 
        "maxTextRecallSize": 5000, 
        "countAndFacetMode": "countRetrievableResults" 
      } 
    }

Configure a query response

When you're setting up the hybrid query, think about the response structure. The response is a flattened rowset. Parameters on the query determine which fields are in each row and how many rows are in the response. The search engine ranks the matching documents and returns the most relevant results.

Fields in a response

Search results are composed of retrievable fields from your search index. A result is either:

All retrievable fields (a REST API default).
Fields explicitly listed in a "select" parameter on the query.

The examples in this article used a "select" statement to specify text (nonvector) fields in the response.

Note

Vectors aren't reverse engineered into human readable text, so avoid returning them in the response. Instead, choose nonvector fields that are representative of the search document. For example, if the query targets a "DescriptionVector" field, return an equivalent text field if you have one ("Description") in the response.

Number of results

A query might match to any number of documents, as many as all of them if the search criteria are weak (for example "search=*" for a null query). Because it's seldom practical to return unbounded results, you should specify a maximum for the overall response:

"top": n results for keyword-only queries (no vector)
"k": n results for vector-only queries
"top": n results for hybrid queries (with or without semantic) that include a "search" parameter

Both "k" and "top" are optional. Unspecified, the default number of results in a response is 50. You can set "top" and "skip" to page through more results or change the default.

Note

If you're using hybrid search in 2024-05-01-preview API, you can control the number of results from the keyword query using maxTextRecallSize. Combine this with a setting for "k" to control the representation from each search subsystem (keyword and vector).

Semantic ranker results

Note

The semantic ranker can take up to 50 results.

If you're using semantic ranker in 2024-05-01-preview API, it's a best practice to set "k" and "maxTextRecallSize" to sum to at least 50 total. You can then restrict the results returned to the user with the "top" parameter.

If you're using semantic ranker in previous APIs do the following:

if doing keyword-only search (no vector) set "top" to 50
if doing hybrid search set "k" to 50, to ensure that the semantic ranker gets at least 50 results.

Ranking

Multiple sets are created for hybrid queries, with or without the optional semantic reranking. Ranking of results is computed by Reciprocal Rank Fusion (RRF).

In this section, compare the responses between single vector search and simple hybrid search for the top result. The different ranking algorithms, HNSW's similarity metric and RRF is this case, produce scores that have different magnitudes. This behavior is by design. RRF scores can appear quite low, even with a high similarity match. Lower scores are a characteristic of the RRF algorithm. In a hybrid query with RRF, more of the reciprocal of the ranked documents are included in the results, given the relatively smaller score of the RRF ranked documents, as opposed to pure vector search.

Single Vector Search: @search.score for results ordered by cosine similarity (default vector similarity distance function).

{
    "@search.score": 0.8399121,
    "HotelId": "49",
    "HotelName": "Swirling Currents Hotel",
    "Description": "Spacious rooms, glamorous suites and residences, rooftop pool, walking access to shopping, dining, entertainment and the city center.",
    "Category": "Luxury",
    "Address": {
    "City": "Arlington"
    }
}

Hybrid Search: @search.score for hybrid results ranked using Reciprocal Rank Fusion.

{
    "@search.score": 0.032786883413791656,
    "HotelId": "49",
    "HotelName": "Swirling Currents Hotel",
    "Description": "Spacious rooms, glamorous suites and residences, rooftop pool, walking access to shopping, dining, entertainment and the city center.",
    "Category": "Luxury",
    "Address": {
    "City": "Arlington"
    }
}

Next steps

As a next step, we recommend reviewing the demo code for Python, C# or JavaScript.

Jaa

Create a hybrid query in Azure AI Search

Prerequisites

Choose an API or tool

Set up a hybrid query in Search Explorer

Hybrid query request (REST API)

Hybrid search with filter

Hybrid search with filters targeting vector subqueries (preview)

Semantic hybrid search

Semantic hybrid search with filter

Set maxTextRecallSize and countAndFacetMode (preview)

Configure a query response

Fields in a response

Number of results

Semantic ranker results

Ranking

Next steps

Palaute

Lisäresursseja