Редактиране

Споделяне чрез

Create an index in Azure AI Search

  • Статия

In this article, learn the steps for defining a schema for a search index and pushing it to a search service. Creating an index establishes the physical data structures on your search service. Once the index exists, load the index as a separate task.

Prerequisites

  • Write permissions as a Search Service Contributor or an admin API key for key-based authentication.

  • An understanding of the data you want to index. A search index is based on external content that you want to make searchable. Searchable content is stored as fields in an index. You should have a clear idea of which source fields you want to make searchable, retrievable, filterable, facetable, and sortable (see the schema checklist for guidance).

  • You must also have a unique field in source data that can be used as the document key (or ID) in the index.

  • A stable index location. Moving an existing index to a different search service isn't supported out-of-the-box. Revisit application requirements and make sure that your existing search service (capacity and location), are sufficient for your needs.

  • Finally, all service tiers have index limits on the number of objects that you can create. For example, if you're experimenting on the Free tier, you can only have three indexes at any given time. Within the index itself, there are limits on vectors and index limits on the number of simple and complex fields.

Document keys

A search index has two requirements: it must have a name and a document key.

A document key is the unique identifier of a search document, and a search document is a collection of fields that completely describes something. For example, if you're indexing a movies data set, a search document contains the title, genre, and duration of a single movie.

In Azure AI Search, a document key must be a string, and it must originate from unique values in the data source that's providing the content to be indexed. A search service doesn't generate key values, but in some scenarios (such as the Azure table indexer) it synthesizes existing values to create a unique key for the documents being indexed.

During incremental indexing, where new and updated content is indexed, incoming documents with new keys are added, while incoming documents with existing keys are either merged or overwritten, depending on whether index fields are null or populated.

Schema checklist

Use this checklist to assist the design decisions for your search index.

  1. Review naming conventions so that index and field names conform to the naming rules.

  2. Review supported data types. The data type affects how the field is used. For example, numeric content is filterable but not full text searchable. The most common data type is Edm.String for searchable text, which is tokenized and queried using the full text search engine. The most common data type for a vector field is Edm.Single but you can use other types as well.

  3. Identify a document key. A document key is an index requirement. It's a single string field and it's populated from a source data field that contains unique values. For example, if you're indexing from Blob Storage, the metadata storage path is often used as the document key because it uniquely identifies each blob in the container.

  4. Identify the fields in your data source that contribute searchable content in the index.

    Searchable nonvector content includes short or long strings that are queried using the full text search engine. If the content is verbose (small phrases or bigger chunks), experiment with different analyzers to see how the text is tokenized.

    Searchable vector content can be images or text (in any language) that exists as a mathematical representation. You can use narrow data types or vector compression to make vector fields smaller.

    Field attribute assignments determine both search behaviors and the physical representation of your index on the search service. Determining how fields should be specified is an iterative process for many customers. To speed up iterations, start with sample data so that you can drop and rebuild easily.

  5. Identify which source fields can be used as filters. Numeric content and short text fields, particularly those with repeating values, are good choices. When working with filters, remember:

    • Filters can be used in vector and nonvector queries, but the filter itself is applied alphanumeric (nonvector) fields in your index.

    • Filterable fields can optionally be used in faceted navigation.

    • Filterable fields are returned in arbitrary order, so consider making them sortable as well.

  6. For vector fields, specify a vector search configuration and the algorithms used for creating navigation paths and filling the embedding space. For more information, see Add vector fields.

    Vector fields have extra properties that nonvector fields don't have, such as which algorithms to use and vector compression.

    Vector fields omit attributes that aren't useful on vector data, such as sorting, filtering, and faceting.

  7. For nonvector fields, determine whether to use the default analyzer ("analyzer": null) or a different analyzer. Analyzers are used to tokenize text fields during indexing and query execution.

    For multi-lingual strings, consider a language analyzer.

    For hyphenated strings or special characters, consider specialized analyzers. One example is keyword that treats the entire contents of a field as a single token. This behavior is useful for data like zip codes, IDs, and some product names. For more information, see Partial term search and patterns with special characters.

Note

Full text search is conducted over terms that are tokenized during indexing. If your queries fail to return the results you expect, test for tokenization to verify the string you're searchin for actually exists. You can try different analyzers on strings to see how tokens are produced for various analyzers.

Create an index

When you're ready to create the index, use a search client that can send the request. You can use the Azure portal or REST APIs for early development and proof-of-concept testing, otherwise it's common to use the Azure SDKs.

During development, plan on frequent rebuilds. Because physical structures are created in the service, dropping and re-creating indexes is necessary for many modifications. You might consider working with a subset of your data to make rebuilds go faster.

Index design through the portal enforces requirements and schema rules for specific data types, such as disallowing full text search capabilities on numeric fields.

  1. Sign in to the Azure portal.

  2. Check for space. Search services are subject to maximum number of indexes, varying by service tier. Make sure you have room for a second index.

  3. In the search service Overview page, choose either option for creating a search index:

    • Add index, an embedded editor for specifying an index schema
    • Import wizards

    The wizard is an end-to-end workflow that creates an indexer, a data source, and a finished index. It also loads the data. If this is more than what you want, use Add index instead.

The following screenshot highlights where Add index and Import data appear on the command bar. After an index is created, you can find it again in the Indexes tab.

Add index command

Tip

After creating an index in the portal, you can copy the JSON representation and add it to your application code.

Set corsOptions for cross-origin queries

Index schemas include a section for setting corsOptions. By default, client-side JavaScript can't call any APIs because browsers prevent all cross-origin requests. To allow cross-origin queries through to your index, enable CORS (Cross-Origin Resource Sharing) by setting the corsOptions attribute. For security reasons, only query APIs support CORS.

"corsOptions": {
  "allowedOrigins": [
    "*"
  ],
  "maxAgeInSeconds": 300

The following properties can be set for CORS:

  • allowedOrigins (required): This is a list of origins that are allowed access to your index. JavaScript code served from these origins is allowed to query your index (assuming the caller provides a valid key or has permissions). Each origin is typically of the form protocol://<fully-qualified-domain-name>:<port> although <port> is often omitted. For more information, see Cross-origin resource sharing (Wikipedia).

    If you want to allow access to all origins, include * as a single item in the allowedOrigins array. This isn't a recommended practice for production search services but it's often useful for development and debugging.

  • maxAgeInSeconds (optional): Browsers use this value to determine the duration (in seconds) to cache CORS preflight responses. This must be a non-negative integer. A longer cache period delivers better performance, but it extends the amount of time a CORS policy needs to take effect. If this value isn't set, a default duration of five minutes is used.

Allowed updates on existing indexes

Create Index creates the physical data structures (files and inverted indexes) on your search service. Once the index is created, your ability to effect changes using Create or Update Index is contingent upon whether your modifications invalidate those physical structures. Most field attributes can't be changed once the field is created in your index.

Alternatively, you can create an index alias that serves as a stable reference in your application code. Instead of updating your code, you can update an index alias to point to newer index versions.

To minimize churn in the design process, the following table describes which elements are fixed and flexible in the schema. Changing a fixed element requires an index rebuild, whereas flexible elements can be changed at any time without impacting the physical implementation.

Element Can be updated?
Name No
Key No
Field names and types No
Field attributes (searchable, filterable, facetable, sortable) No
Field attribute (retrievable) Yes
Stored (applies to vectors) No
Analyzer You can add and modify custom analyzers in the index. Regarding analyzer assignments on string fields, you can only modify searchAnalyzer. All other assignments and modifications require a rebuild.
Scoring profiles Yes
Suggesters No
cross-origin resource sharing (CORS) Yes
Encryption Yes

Next steps

Use the following links to become familiar with loading an index with data, or extending an index with a synonyms map.