Model complex data types in Azure AI Search

External datasets used to populate an Azure AI Search index can come in many shapes. Sometimes they include hierarchical or nested substructures. Examples might include multiple addresses for a single customer, multiple colors and sizes for a single product, multiple authors of a single book, and so on. In modeling terms, you might see these structures referred to as complex, compound, composite, or aggregate data types. The term Azure AI Search uses for this concept is complex type. In Azure AI Search, complex types are modeled using complex fields. A complex field is a field that contains children (subfields) which can be of any data type, including other complex types. This works in a similar way as structured data types in a programming language.

Complex fields represent either a single object in the document, or an array of objects, depending on the data type. Fields of type Edm.ComplexType represent single objects, while fields of type Collection(Edm.ComplexType) represent arrays of objects.

Azure AI Search natively supports complex types and collections. These types allow you to model almost any JSON structure in an Azure AI Search index. In previous versions of Azure AI Search APIs, only flattened row sets could be imported. In the newest version, your index can now more closely correspond to source data. In other words, if your source data has complex types, your index can have complex types also.

To get started, we recommend the Hotels data set, which you can load in the Import data wizard in the Azure portal. The wizard detects complex types in the source and suggests an index schema based on the detected structures.

Note

Support for complex types became generally available starting in api-version=2019-05-06.

If your search solution is built on earlier workarounds of flattened datasets in a collection, you should change your index to include complex types as supported in the newest API version. For more information about upgrading API versions, see Upgrade to the newest REST API version or Upgrade to the newest .NET SDK version.

Example of a complex structure

The following JSON document is composed of simple fields and complex fields. Complex fields, such as Address and Rooms, have subfields. Address has a single set of values for those subfields, since it's a single object in the document. In contrast, Rooms has multiple sets of values for its subfields, one for each object in the collection.

{
  "HotelId": "1",
  "HotelName": "Stay-Kay City Hotel",
  "Description": "Ideally located on the main commercial artery of the city in the heart of New York.",
  "Tags": ["Free wifi", "on-site parking", "indoor pool", "continental breakfast"],
  "Address": {
    "StreetAddress": "677 5th Ave",
    "City": "New York",
    "StateProvince": "NY"
  },
  "Rooms": [
    {
      "Description": "Budget Room, 1 Queen Bed (Cityside)",
      "RoomNumber": 1105,
      "BaseRate": 96.99,
    },
    {
      "Description": "Deluxe Room, 2 Double Beds (City View)",
      "Type": "Deluxe Room",
      "BaseRate": 150.99,
    }
    . . .
  ]
}

Create complex fields

As with any index definition, you can use the portal, REST API, or .NET SDK to create a schema that includes complex types.

Other Azure SDKs provide samples in Python, Java, and JavaScript.

  1. Sign in to the Azure portal.

  2. On the search service Overview page, select the Indexes tab.

  3. Open an existing index or create a new index.

  4. Select the Fields tab, and then select Add field. An empty field is added. If you're working with an existing fields collection, scroll down to set up the field.

  5. Give the field a name and set the type to either Edm.ComplexType or Collection(Edm.ComplexType).

  6. Select the ellipses on the far right, and then select either Add field or Add subfield, and then assign attributes.

Complex collection limits

During indexing, you can have a maximum of 3,000 elements across all complex collections within a single document. An element of a complex collection is a member of that collection. For Rooms (the only complex collection in the Hotel example), each room is an element. In the example above, if the "Stay-Kay City Hotel" had 500 rooms, the hotel document would have 500 room elements. For nested complex collections, each nested element is also counted, in addition to the outer (parent) element.

This limit applies only to complex collections, and not complex types (like Address) or string collections (like Tags).

Update complex fields

All of the reindexing rules that apply to fields in general still apply to complex fields. Adding a new field to a complex type doesn't require an index rebuild, but most other modifications do require a rebuild.

Structural updates to the definition

You can add new subfields to a complex field at any time without the need for an index rebuild. For example, adding "ZipCode" to Address or "Amenities" to Rooms is allowed, just like adding a top-level field to an index. Existing documents have a null value for new fields until you explicitly populate those fields by updating your data.

Notice that within a complex type, each subfield has a type and can have attributes, just as top-level fields do

Data updates

Updating existing documents in an index with the upload action works the same way for complex and simple fields: all fields are replaced. However, merge (or mergeOrUpload when applied to an existing document) doesn't work the same across all fields. Specifically, merge doesn't support merging elements within a collection. This limitation exists for collections of primitive types and complex collections. To update a collection, you need to retrieve the full collection value, make changes, and then include the new collection in the Index API request.

Search complex fields in text queries

Free-form search expressions work as expected with complex types. If any searchable field or subfield anywhere in a document matches, then the document itself is a match.

Queries get more nuanced when you have multiple terms and operators, and some terms have field names specified, as is possible with the Lucene syntax. For example, this query attempts to match two terms, "Portland" and "OR", against two subfields of the Address field:

search=Address/City:Portland AND Address/State:OR

Queries like this are uncorrelated for full-text search, unlike filters. In filters, queries over subfields of a complex collection are correlated using range variables in any or all. The Lucene query above returns documents containing both "Portland, Maine" and "Portland, Oregon", along with other cities in Oregon. This happens because each clause applies to all values of its field in the entire document, so there's no concept of a "current subdocument". For more information on this, see Understanding OData collection filters in Azure AI Search.

Search complex fields in RAG queries

A RAG pattern passes search results to a chat model for generative AI and conversational search. By default, search results passed to an LLM are a flattened rowset. However, if your index has complex types, your query can provide those fields if you first convert the search results to JSON, and then pass the JSON to the LLM.

A partial example illustrates the technique:

  • Indicate the fields you want in the prompt or in the query
  • Make sure the fields are searchable and retrievable in the index
  • Select the fields for the search results
  • Format the results as JSON
  • Send the request for chat completion to the model provider
import json

# Query is the question being asked. It's sent to the search engine and the LLM.
query="Can you recommend a few hotels that offer complimentary breakfast? Tell me their description, address, tags, and the rate for one room they have which sleep 4 people."

# Set up the search results and the chat thread.
# Retrieve the selected fields from the search index related to the question.
selected_fields = ["HotelName","Description","Address","Rooms","Tags"]
search_results = search_client.search(
    search_text=query,
    top=5,
    select=selected_fields,
    query_type="semantic"
)
sources_filtered = [{field: result[field] for field in selected_fields} for result in search_results]
sources_formatted = "\n".join([json.dumps(source) for source in sources_filtered])

response = openai_client.chat.completions.create(
    messages=[
        {
            "role": "user",
            "content": GROUNDED_PROMPT.format(query=query, sources=sources_formatted)
        }
    ],
    model=AZURE_DEPLOYMENT_MODEL
)

print(response.choices[0].message.content)

For the end-to-end example, see Quickstart: Generative search (RAG) with grounding data from Azure AI Search.

Select complex fields

The $select parameter is used to choose which fields are returned in search results. To use this parameter to select specific subfields of a complex field, include the parent field and subfield separated by a slash (/).

$select=HotelName, Address/City, Rooms/BaseRate

Fields must be marked as Retrievable in the index if you want them in search results. Only fields marked as Retrievable can be used in a $select statement.

Filter, facet, and sort complex fields

The same OData path syntax used for filtering and fielded searches can also be used for faceting, sorting, and selecting fields in a search request. For complex types, rules apply that govern which subfields can be marked as sortable or facetable. For more information on these rules, see the Create Index API reference.

Faceting subfields

Any subfield can be marked as facetable unless it is of type Edm.GeographyPoint or Collection(Edm.GeographyPoint).

The document counts returned in the facet results are calculated for the parent document (a hotel), not the subdocuments in a complex collection (rooms). For example, suppose a hotel has 20 rooms of type "suite". Given this facet parameter facet=Rooms/Type, the facet count is one for the hotel, not 20 for the rooms.

Sorting complex fields

Sort operations apply to documents (Hotels) and not subdocuments (Rooms). When you have a complex type collection, such as Rooms, it's important to realize that you can't sort on Rooms at all. In fact, you can't sort on any collection.

Sort operations work when fields have a single value per document, whether the field is a simple field, or a subfield in a complex type. For example, Address/City is allowed to be sortable because there's only one address per hotel, so $orderby=Address/City sorts hotels by city.

Filtering on complex fields

You can refer to subfields of a complex field in a filter expression. Just use the same OData path syntax that's used for faceting, sorting, and selecting fields. For example, the following filter returns all hotels in Canada:

$filter=Address/Country eq 'Canada'

To filter on a complex collection field, you can use a lambda expression with the any and all operators. In that case, the range variable of the lambda expression is an object with subfields. You can refer to those subfields with the standard OData path syntax. For example, the following filter returns all hotels with at least one deluxe room and all nonsmoking rooms:

$filter=Rooms/any(room: room/Type eq 'Deluxe Room') and Rooms/all(room: not room/SmokingAllowed)

As with top-level simple fields, simple subfields of complex fields can only be included in filters if they have the filterable attribute set to true in the index definition. For more information, see the Create Index API reference.

Workaround for the complex collection limit

Recall that Azure AI Search limits complex objects in a collection to 3,000 objects per document. Exceeding this limit results in the following message:

A collection in your document exceeds the maximum elements across all complex collections limit. 
The document with key '1052' has '4303' objects in collections (JSON arrays). 
At most '3000' objects are allowed to be in collections across the entire document. 
Remove objects from collections and try indexing the document again."

If you need more than 3,000 items, you can pipe (|) or use any form of delimiter to delimit the values, concatenate them, and store them as a delimited string. There's no limitation on the number of strings stored in an array. Storing complex values as strings bypasses the complex collection limitation.

To illustrate, assume you have a "searchScope" array with more than 3,000 elements:


"searchScope": [
  {
     "countryCode": "FRA",
     "productCode": 1234,
     "categoryCode": "C100" 
  },
  {
     "countryCode": "USA",
     "productCode": 1235,
     "categoryCode": "C200" 
  }
  . . .
]

The workaround for storing the values as a delimited string might look like this:

"searchScope": [
        "|FRA|1234|C100|",
        "|FRA|*|*|",
        "|*|1234|*|",
        "|*|*|C100|",
        "|FRA|*|C100|",
        "|*|1234|C100|"
]

Storing all of the search variants in the delimited string is helpful in search scenarios where you want to search for items that have just "FRA" or "1234" or another combination within the array.

Here's a filter formatting snippet in C# that converts inputs into searchable strings:

foreach (var filterItem in filterCombinations)
        {
            var formattedCondition = $"searchScope/any(s: s eq '{filterItem}')";
            combFilter.Append(combFilter.Length > 0 ? " or (" + formattedCondition + ")" : "(" + formattedCondition + ")");
        }

The following list provides inputs and search strings (outputs) side by side:

  • For "FRA" county code and the "1234" product code, the formatted output is |FRA|1234|*|.

  • For "1234" product code, the formatted output is |*|1234|*|.

  • For "C100" category code, the formatted output is |*|*|C100|.

Only provide the wildcard (*) if you're implementing the string array workaround. Otherwise, if you're using a complex type, your filter might look like this example:

var countryFilter = $"searchScope/any(ss: search.in(countryCode ,'FRA'))";
var catgFilter = $"searchScope/any(ss: search.in(categoryCode ,'C100'))";
var combinedCountryCategoryFilter = "(" + countryFilter + " and " + catgFilter + ")";

If you implement the workaround, be sure to test extentively.

Next steps

Try the Hotels data set in the Import data wizard. You need the Azure Cosmos DB connection information provided in the readme to access the data.

With that information in hand, your first step in the wizard is to create a new Azure Cosmos DB data source. Further on in the wizard, when you get to the target index page, you see an index with complex types. Create and load this index, and then execute queries to understand the new structure.