Upgrade to the latest REST API in Azure AI Search

Стаття
10/01/2024

Use this article to migrate data plane calls to newer versions of the Search REST APIs.

2024-07-01 is the most recent stable version.
2024-09-01-preview is the most recent preview API version.

Upgrade instructions focus on code changes that get you through breaking changes from previous versions so that existing code runs the same as before, but on the newer API version. Once your code is in working order, you can decide whether to adopt newer features. To learn more about new features, see vector code samples and What's New.

We recommend upgrading API versions in succession, working through each version until you get to the newest one.

2023-07-01-preview was the first REST API for vector support. Do not use this API version. It's now deprecated and you should migrate to either stable or newer preview REST APIs immediately.

Note

REST API reference docs are now versioned. For version-specific content, open a reference page and then use the selector located to the right, above the table of contents, to pick your version.

When to upgrade

Azure AI Search breaks backward compatibility as a last resort. Upgrade is necessary when:

Your code references a retired or unsupported API version and is subject to one or more breaking changes. You must address breaking changes if your code targets 2023-07-10-preview for vectors, 2020-06-01-preview for semantic ranker, and 2019-05-06 for obsolete skills and workarounds.
Your code fails when unrecognized properties are returned in an API response. As a best practice, your application should ignore properties that it doesn't understand.
Your code persists API requests and tries to resend them to the new API version. For example, this might happen if your application persists continuation tokens returned from the Search API (for more information, look for @search.nextPageParameters in the Search API Reference).

How to upgrade

Review the release notes for each API version.
Update the api-version parameter, specified in the request header, to a newer version.

In your application code that makes direct calls to the REST APIs, search for all instances of the existing version and then replace it with the new version. For more information about structuring a REST call, see Quickstart: using REST.

If you're using an Azure SDK, those packages target specific versions of the REST API. Package updates might coincide with a REST API update, but each SDK is on its own release schedule that ships independently of Azure AI Search REST API versions. Check the change log of your SDK package to determine whether a package release targets the latest REST API version.
Review the breaking changes documented in this article and implement the workarounds. Start with the version used by your code and resolve any breaking change for each newer API version until you get to the newest stable or preview release.

Breaking change for client code that reads connection information

Effective March 29, 2024 and applicable to all supported REST APIs:

GET Skillset, GET Index, and GET Indexer no longer return keys or connection properties in a response. This is a breaking change if you have downstream code that reads keys or connections (sensitive data) from a GET response.
If you need to retrieve admin or query API keys for your search service, use the Management REST APIs.
If you need to retrieve connection strings of another Azure resource such as Azure Storage or Azure Cosmos DB, use the APIs of that resource and published guidance to obtain the information.

Breaking change for semantic ranker

Semantic ranker is generally available in 2023-11-01. Here are the breaking changes for semantic ranker from earlier releases:

In all versions after 2020-06-01-preview: semanticConfiguration replaces searchFields as the mechanism for specifying which fields to use for L2 ranking.
For all API versions, updates on July 14, 2023 to the Microsoft-hosted semantic models made semantic ranker language-agnostic, effectively decommissioning the queryLanguage property. There's no "breaking change" in code, but the property is ignored.

See Migrate from preview version to transition your code to use semanticConfiguration.

Upgrade to 2024-09-01-preview

2024-09-01-preview adds Matryoshka Representation Learning (MRL) compression for text-embedding-3 models, targeted vector filtering for hybrid queries, vector subscore details for debugging, and token chunking for Text Split skill.

If you're upgrading from 2024-05-01-preview, you can use the new preview APIs with no change to existing code.

Upgrade to 2024-07-01

2024-07-01 is a general release. The former preview features are now generally available: integrated chunking and vectorization (Text Split skill, AzureOpenAIEmbedding skill), query vectorizer based on AzureOpenAIEmbedding, vector compression (scalar quantization, binary quantization, stored property, narrow data types).

There are no breaking changes if you upgrade from 2024-05-01-preview to stable. To use the new stable release, change the API version and test your code.

There are breaking changes if you upgrade directly from 2023-11-01. Follow the steps outlined for each newer preview to migrate from 2023-11-01 to 2024-07-01.

Upgrade to 2024-05-01-preview

2024-05-01-preview adds OneLake index, support for binary vectors, and support for more embedding models.

If you're upgrading from 2024-03-01-preview, the AzureOpenAIEmbedding skill now requires a model name and dimensions property.

Search your codebase for AzureOpenAIEmbedding references.
Set modelName to "text-embedding-ada-002" and set dimensions to "1536".

Upgrade to 2024-03-01-preview

2024-03-01-preview adds narrow data types, scalar quantization, and vector storage options.

If you're upgrading from 2023-10-01-preview, there are no breaking changes. However, there's one behavior difference: for 2023-11-01 and newer previews, the vectorFilterMode default changed from postfilter to prefilter for filter expressions.

Search your codebase for vectorFilterMode references.
If the property is explicitly set, no action is required. If you used the default, be aware that the new default behavior is to filter before query execution. If you want post-query filtering, explicitly set vectorFilterMode to postfilter to retain the old behavior.

Upgrade to 2023-11-01

2023-11-01 is a general release. The former preview features are now generally available: semantic ranker, vector index and query support.

There are no breaking changes from 2023-10-01-preview, but there are multiple breaking changes from 2023-07-01-preview to 2023-11-01. For more information, see Upgrade from 2023-07-01-preview.

To use the new stable release, change the API version and test your code.

Upgrade to 2023-10-01-preview

2023-10-01-preview was the first preview version to add built-in data chunking and vectorization during indexing and built-in query vectorization. It also supports vector indexing and queries from the previous version.

If you're upgrading from the previous version, the next section has the steps.

Upgrade from 2023-07-01-preview

Do not use this API version. It implements a vector query syntax that's incompatible with any newer API version.

2023-07-01-preview is now deprecated, so you shouldn't base new code on this version, nor should you upgrade to this version under any circumstances. This section explains the migration path from 2023-07-01-preview to any newer API version.

Portal upgrade for vector indexes

Azure portal supports a one-click upgrade path for 2023-07-01-preview indexes. It detects vector fields and provides a Migrate button.

Migration path is from 2023-07-01-preview to 2024-05-01-preview.
Updates are limited to vector field definitions and vector search algorithm configurations.
Updates are one-way. You can't reverse the upgrade. Once the index is upgraded, you must use 2024-05-01-preview or later to query the index.

There's no portal migration for upgrading vector query syntax. See code upgrades for query syntax changes.

Before selecting Migrate, select Edit JSON to review the updated schema first. You should find a schema that conforms to the changes described in the code upgrade section. Portal migration only handles indexes with one vector search algorithm configuration. It creates a default profile that maps to the 2023-07-01-preview vector search algorithm. Indexes with multiple vector search configurations require manual migration.

Code upgrade for vector indexes and queries

Vector search support was introduced in Create or Update Index (2023-07-01-preview).

Upgrading from 2023-07-01-preview to any newer stable or preview version requires:

Renaming and restructuring the vector configuration in the index
Rewriting your vector queries

Use the instructions in this section to migrate vector fields, configuration, and queries from 2023-07-01-preview.

Call Get Index to retrieve the existing definition.

Modify the vector search configuration. 2023-11-01 and later versions introduce the concept of vector profiles that bundle vector-related configurations under one name. Newer versions also rename algorithmConfigurations to algorithms.

Rename algorithmConfigurations to algorithms. This is only a renaming of the array. The contents are backwards compatible. This means your existing HNSW configuration parameters can be used.
Add profiles, giving a name and an algorithm configuration for each one.

Before migration (2023-07-01-preview):

  "vectorSearch": {
    "algorithmConfigurations": [
        {
            "name": "myHnswConfig",
            "kind": "hnsw",
            "hnswParameters": {
                "m": 4,
                "efConstruction": 400,
                "efSearch": 500,
                "metric": "cosine"
            }
        }
    ]}

After migration (2023-11-01):

  "vectorSearch": {
    "algorithms": [
      {
        "name": "myHnswConfig",
        "kind": "hnsw",
        "hnswParameters": {
          "m": 4,
          "efConstruction": 400,
          "efSearch": 500,
          "metric": "cosine"
        }
      }
    ],
    "profiles": [
      {
        "name": "myHnswProfile",
        "algorithm": "myHnswConfig"
      }
    ]
  }

Modify vector field definitions, replacing vectorSearchConfiguration with vectorSearchProfile. Make sure the profile name resolves to a new vector profile definition, and not the algorithm configuration name. Other vector field properties remain unchanged. For example, they can't be filterable, sortable, or facetable, nor use analyzers or normalizers or synonym maps.

Before (2023-07-01-preview):

  {
      "name": "contentVector",
      "type": "Collection(Edm.Single)",
      "key": false,
      "searchable": true,
      "retrievable": true,
      "filterable": false,  
      "sortable": false,  
      "facetable": false,
      "analyzer": "",
      "searchAnalyzer": "",
      "indexAnalyzer": "",
      "normalizer": "",
      "synonymMaps": "", 
      "dimensions": 1536,
      "vectorSearchConfiguration": "myHnswConfig"
  }

After (2023-11-01):

  {
    "name": "contentVector",
    "type": "Collection(Edm.Single)",
    "searchable": true,
    "retrievable": true,
    "filterable": false,  
    "sortable": false,  
    "facetable": false,
    "analyzer": "",
    "searchAnalyzer": "",
    "indexAnalyzer": "",
    "normalizer": "",
    "synonymMaps": "", 
    "dimensions": 1536,
    "vectorSearchProfile": "myHnswProfile"
  }

Call Create or Update Index to post the changes.

Modify Search POST to change the query syntax. This API change enables support for polymorphic vector query types.

Rename vectors to vectorQueries.
For each vector query, add kind, setting it to vector.
For each vector query, rename value to vector.
Optionally, add vectorFilterMode if you're using filter expressions. The default is prefilter for indexes created after 2023-10-01. Indexes created before that date only support postfilter, regardless of how you set the filter mode.

Before (2023-07-01-preview):

{
    "search": (this parameter is ignored in vector search),
    "vectors": [
      {
        "value": [
            0.103,
            0.0712,
            0.0852,
            0.1547,
            0.1183
        ],
        "fields": "contentVector",
        "k": 5
      }
    ],
    "select": "title, content, category"
}

After (2023-11-01):

{
  "search": "(this parameter is ignored in vector search)",
  "vectorQueries": [
    {
      "kind": "vector",
      "vector": [
        0.103,
        0.0712,
        0.0852,
        0.1547,
        0.1183
      ],
      "fields": "contentVector",
      "k": 5
    }
  ],
  "vectorFilterMode": "preFilter",
  "select": "title, content, category"
}

These steps complete the migration to 2023-11-01 stable API version or newer preview API versions.

Upgrade to 2020-06-30

In this version, there's one breaking change and several behavioral differences. Generally available features include:

Knowledge store, persistent storage of enriched content created through skillsets, created for downstream analysis and processing through other applications. A knowledge store is created through Azure AI Search REST APIs but it resides in Azure Storage.

Breaking change

Code written against earlier API versions breaks on 2020-06-30 and later if code contains the following functionality:

Any Edm.Date literals (a date composed of year-month-day, such as 2020-12-12) in filter expressions must follow the Edm.DateTimeOffset format: 2020-12-12T00:00:00Z. This change was necessary to handle erroneous or unexpected query results due to timezone differences.

Behavior changes

BM25 ranking algorithm replaces the previous ranking algorithm with newer technology. Services created after 2019 use this algorithm automatically. For older services, you must set parameters to use the new algorithm.
Ordered results for null values have changed in this version, with null values appearing first if the sort is asc and last if the sort is desc. If you wrote code to handle how null values are sorted, be aware of this change.

Upgrade to 2019-05-06

Features that became generally available in this API version include:

Autocomplete is a typeahead feature that completes a partially specified term input.
Complex types provides native support for structured object data in search index.
JsonLines parsing modes, part of Azure Blob indexing, creates one search document per JSON entity that is separated by a newline.
AI enrichment provides indexing that uses the AI enrichment engines of Azure AI services.

Breaking changes

Code written against an earlier API version breaks on 2019-05-06 and later if it contains the following functionality:

Type property for Azure Cosmos DB. For indexers targeting an Azure Cosmos DB for NoSQL API data source, change "type": "documentdb" to "type": "cosmosdb".
If your indexer error handling includes references to the status property, you should remove it. We removed status from the error response because it wasn't providing useful information.
Data source connection strings are no longer returned in the response. From API versions 2019-05-06 and 2019-05-06-Preview onwards, the data source API no longer returns connection strings in the response of any REST operation. In previous API versions, for data sources created using POST, Azure AI Search returned 201 followed by the OData response, which contained the connection string in plain text.
Named Entity Recognition cognitive skill is retired. If you called the Name Entity Recognition skill in your code, the call fails. Replacement functionality is Entity Recognition Skill (V3). Follow the recommendations in Deprecated skills to migrate to a supported skill.

Upgrading complex types

API version 2019-05-06 added formal support for complex types. If your code implemented previous recommendations for complex type equivalency in 2017-11-11-Preview or 2016-09-01-Preview, there are some new and changed limits starting in version 2019-05-06 of which you need to be aware:

The limits on the depth of subfields and the number of complex collections per index have been lowered. If you created indexes that exceed these limits using the preview api-versions, any attempt to update or recreate them using API version 2019-05-06 will fail. If you find yourself in this situation, you need to redesign your schema to fit within the new limits and then rebuild your index.
There's a new limit starting in api-version 2019-05-06 on the number of elements of complex collections per document. If you created indexes with documents that exceed these limits using the preview api-versions, any attempt to reindex that data using api-version 2019-05-06 fails. If you find yourself in this situation, you need to reduce the number of complex collection elements per document before reindexing your data.

For more information, see Service limits for Azure AI Search.

How to upgrade an old complex type structure

If your code is using complex types with one of the older preview API versions, you might be using an index definition format that looks like this:

{
  "name": "hotels",  
  "fields": [
    { "name": "HotelId", "type": "Edm.String", "key": true, "filterable": true },
    { "name": "HotelName", "type": "Edm.String", "searchable": true, "filterable": false, "sortable": true, "facetable": false },
    { "name": "Description", "type": "Edm.String", "searchable": true, "filterable": false, "sortable": false, "facetable": false, "analyzer": "en.microsoft" },
    { "name": "Description_fr", "type": "Edm.String", "searchable": true, "filterable": false, "sortable": false, "facetable": false, "analyzer": "fr.microsoft" },
    { "name": "Category", "type": "Edm.String", "searchable": true, "filterable": true, "sortable": true, "facetable": true },
    { "name": "Tags", "type": "Collection(Edm.String)", "searchable": true, "filterable": true, "sortable": false, "facetable": true, "analyzer": "tagsAnalyzer" },
    { "name": "ParkingIncluded", "type": "Edm.Boolean", "filterable": true, "sortable": true, "facetable": true },
    { "name": "LastRenovationDate", "type": "Edm.DateTimeOffset", "filterable": true, "sortable": true, "facetable": true },
    { "name": "Rating", "type": "Edm.Double", "filterable": true, "sortable": true, "facetable": true },
    { "name": "Address", "type": "Edm.ComplexType" },
    { "name": "Address/StreetAddress", "type": "Edm.String", "filterable": false, "sortable": false, "facetable": false, "searchable": true },
    { "name": "Address/City", "type": "Edm.String", "searchable": true, "filterable": true, "sortable": true, "facetable": true },
    { "name": "Address/StateProvince", "type": "Edm.String", "searchable": true, "filterable": true, "sortable": true, "facetable": true },
    { "name": "Address/PostalCode", "type": "Edm.String", "searchable": true, "filterable": true, "sortable": true, "facetable": true },
    { "name": "Address/Country", "type": "Edm.String", "searchable": true, "filterable": true, "sortable": true, "facetable": true },
    { "name": "Location", "type": "Edm.GeographyPoint", "filterable": true, "sortable": true },
    { "name": "Rooms", "type": "Collection(Edm.ComplexType)" }, 
    { "name": "Rooms/Description", "type": "Edm.String", "searchable": true, "filterable": false, "sortable": false, "facetable": false, "analyzer": "en.lucene" },
    { "name": "Rooms/Description_fr", "type": "Edm.String", "searchable": true, "filterable": false, "sortable": false, "facetable": false, "analyzer": "fr.lucene" },
    { "name": "Rooms/Type", "type": "Edm.String", "searchable": true },
    { "name": "Rooms/BaseRate", "type": "Edm.Double", "filterable": true, "facetable": true },
    { "name": "Rooms/BedOptions", "type": "Edm.String", "searchable": true },
    { "name": "Rooms/SleepsCount", "type": "Edm.Int32", "filterable": true, "facetable": true },
    { "name": "Rooms/SmokingAllowed", "type": "Edm.Boolean", "filterable": true, "facetable": true },
    { "name": "Rooms/Tags", "type": "Collection(Edm.String)", "searchable": true, "filterable": true, "facetable": true, "analyzer": "tagsAnalyzer" }
  ]
}

A newer tree-like format for defining index fields was introduced in API version 2017-11-11-Preview. In the new format, each complex field has a fields collection where its subfields are defined. In API version 2019-05-06, this new format is used exclusively and attempting to create or update an index using the old format will fail. If you have indexes created using the old format, you'll need to use API version 2017-11-11-Preview to update them to the new format before they can be managed using API version 2019-05-06.

You can update flat indexes to the new format with the following steps using API version 2017-11-11-Preview:

Perform a GET request to retrieve your index. If it’s already in the new format, you’re done.
Translate the index from the flat format to the new format. You have to write code for this task since there's no sample code available at the time of this writing.
Perform a PUT request to update the index to the new format. Avoid changing any other details of the index, such as the searchability/filterability of fields, because changes that affect the physical expression of existing index isn't allowed by the Update Index API.