แก้ไข

แชร์ผ่าน


Index JSON blobs and files in Azure AI Search

Applies to: Blob indexers, File indexers

For blob indexing in Azure AI Search, this article shows you how to set properties for blobs or files consisting of JSON documents. JSON files in Azure Blob Storage or Azure Files commonly assume any of these forms:

  • A single JSON document
  • A JSON document containing an array of well-formed JSON elements
  • A JSON document containing multiple entities, separated by a newline

The blob indexer provides a parsingMode parameter to optimize the output of the search document based on JSON structure. Parsing modes consist of the following options:

parsingMode JSON document Description
json One per blob (default) Parses JSON blobs as a single chunk of text. Each JSON blob becomes a single search document.
jsonArray Multiple per blob Parses a JSON array in the blob, where each element of the array becomes a separate search document.
jsonLines Multiple per blob Parses a blob that contains multiple JSON entities (also an array), with individual elements separated by a newline. The indexer starts a new search document after each new line.

For both jsonArray and jsonLines, you should review Indexing one blob to produce many search documents to understand how the blob indexer handles disambiguation of the document key for multiple search documents produced from the same blob.

Within the indexer definition, you can optionally set field mappings to choose which properties of the source JSON document are used to populate your target search index. For example, when using the jsonArray parsing mode, if the array exists as a lower-level property, you can set a "documentRoot" property indicating where the array is placed within the blob.

Note

When a JSON parsing mode is used, Azure AI Search assumes that all blobs use the same parser (either for json, jsonArray or jsonLines). If you have a mix of different file types in the same data source, consider using file extension filters to control which files are imported.

The following sections describe each mode in more detail. If you're unfamiliar with indexer clients and concepts, see Create a search indexer. You should also be familiar with the details of basic blob indexer configuration, which isn't repeated here.

Index single JSON documents (one per blob)

By default, blob indexers parse JSON blobs as a single chunk of text, one search document for each blob in a container. If the JSON is structured, the search document can reflect that structure, with individual elements represented as individual fields. For example, assume you have the following JSON document in Azure Blob Storage:

{
    "article" : {
        "text" : "A hopefully useful article explaining how to parse JSON blobs",
        "datePublished" : "2020-04-13",
        "tags" : [ "search", "storage", "howto" ]    
    }
}

The blob indexer parses the JSON document into a single search document, loading an index by matching "text", "datePublished", and "tags" from the source against identically named and typed target index fields. Given an index with "text", "datePublished, and "tags" fields, the blob indexer can infer the correct mapping without a field mapping present in the request.

Although the default behavior is one search document per JSON blob, setting the json parsing mode changes the internal field mappings for content, promoting fields inside content to actual fields in the search index. An example indexer definition for the json parsing mode might look like this:

POST https://[service name].search.windows.net/indexers?api-version=2024-07-01
Content-Type: application/json
api-key: [admin key]

{
    "name" : "my-json-indexer",
    "dataSourceName" : "my-blob-datasource",
    "targetIndexName" : "my-target-index",
    "parameters" : { "configuration" : { "parsingMode" : "json" } }
}

Note

As with all indexers, if fields do not clearly match, you should expect to explicitly specify individual field mappings unless you are using the implicit fields mappings available for blob content and metadata, as described in basic blob indexer configuration.

json example (single hotel JSON files)

The hotel JSON document data set on GitHub is helpful for testing JSON parsing, where each blob represents a structured JSON file. You can upload the data files to Blob Storage and use the Import data wizard to quickly evaluate how this content is parsed into individual search documents.

The data set consists of five blobs, each containing a hotel document with an address collection and a rooms collection. The blob indexer detects both collections and reflects the structure of the input documents in the index schema.

Parse JSON arrays

Alternatively, you can use the JSON array option. This option is useful when blobs contain an array of well-formed JSON objects, and you want each element to become a separate search document. Using jsonArrays, the following JSON blob produces three separate documents, each with "id" and "text" fields.

[
    { "id" : "1", "text" : "example 1" },
    { "id" : "2", "text" : "example 2" },
    { "id" : "3", "text" : "example 3" }
]

The parameters property on the indexer contains parsing mode values. For a JSON array, the indexer definition should look similar to the following example.

POST https://[service name].search.windows.net/indexers?api-version=2024-07-01
Content-Type: application/json
api-key: [admin key]

{
    "name" : "my-json-indexer",
    "dataSourceName" : "my-blob-datasource",
    "targetIndexName" : "my-target-index",
    "parameters" : { "configuration" : { "parsingMode" : "jsonArray" } }
}

jsonArrays example

The New York Philharmonic JSON data set on GitHub is helpful for testing JSON array parsing. You can upload the data files to Blob storage and use the Import data wizard to quickly evaluate how this content is parsed into individual search documents.

The data set consists of eight blobs, each containing a JSON array of entities, for a total of 100 entities. The entities vary as to which fields are populated, but the end result is one search document per entity, from all arrays, in all blobs.

Parsing nested JSON arrays

For JSON arrays having nested elements, you can specify a documentRoot to indicate a multi-level structure. For example, if your blobs look like this:

{
    "level1" : {
        "level2" : [
            { "id" : "1", "text" : "Use the documentRoot property" },
            { "id" : "2", "text" : "to pluck the array you want to index" },
            { "id" : "3", "text" : "even if it's nested inside the document" }  
        ]
    }
}

Use this configuration to index the array contained in the level2 property:

{
    "name" : "my-json-array-indexer",
    ... other indexer properties
    "parameters" : { "configuration" : { "parsingMode" : "jsonArray", "documentRoot" : "/level1/level2" } }
}

Parse JSON entities separated by newlines

If your blob contains multiple JSON entities separated by a newline, and you want each element to become a separate search document, use jsonLines.

{ "id" : "1", "text" : "example 1" }
{ "id" : "2", "text" : "example 2" }
{ "id" : "3", "text" : "example 3" }

For JSON lines, the indexer definition should look similar to the following example.

POST https://[service name].search.windows.net/indexers?api-version=2024-07-01
Content-Type: application/json
api-key: [admin key]

{
    "name" : "my-json-indexer",
    "dataSourceName" : "my-blob-datasource",
    "targetIndexName" : "my-target-index",
    "parameters" : { "configuration" : { "parsingMode" : "jsonLines" } }
}

Map JSON fields to search fields

Field mappings associate a source field with a destination field in situations where the field names and types aren't identical. But field mappings can also be used to match parts of a JSON document and "lift" them into top-level fields of the search document.

The following example illustrates this scenario. For more information about field mappings in general, see field mappings.

{
    "article" : {
        "text" : "A hopefully useful article explaining how to parse JSON blobs",
        "datePublished" : "2016-04-13"
        "tags" : [ "search", "storage", "howto" ]    
    }
}

Assume a search index with the following fields: text of type Edm.String, date of type Edm.DateTimeOffset, and tags of type Collection(Edm.String). Notice the discrepancy between "datePublished" in the source and date field in the index. To map your JSON into the desired shape, use the following field mappings:

"fieldMappings" : [
    { "sourceFieldName" : "/article/text", "targetFieldName" : "text" },
    { "sourceFieldName" : "/article/datePublished", "targetFieldName" : "date" },
    { "sourceFieldName" : "/article/tags", "targetFieldName" : "tags" }
    ]

Source fields are specified using the JSON Pointer notation. You start with a forward slash to refer to the root of your JSON document, then pick the desired property (at arbitrary level of nesting) by using forward slash-separated path.

You can also refer to individual array elements by using a zero-based index. For example, to pick the first element of the "tags" array from the above example, use a field mapping like this:

{ "sourceFieldName" : "/article/tags/0", "targetFieldName" : "firstTag" }

Note

If "sourceFieldName" refers to a property that doesn't exist in the JSON blob, that mapping is skipped without an error. This behavior allows indexing to continue for JSON blobs that have a different schema (which is a common use case). Because there is no validation check, check the mappings carefully for typos so that you aren't losing documents for the wrong reason.

Next steps