Geolocation_Data

Geolocation Data

Labeled geolocation data

In case the source data is structured with a well-defined schema, one can add geolocation data from source documents to the index to be able to query it from Azure search. Here's a video showing the possibilities of searching for entries near a given location or within a given search radius.

Create or add a field to your index with data type Edm.GeographyPoint (see this link for supported data types).

As an example, the fields collection could contain an item:

{
    "name": "my_index_geolocation",
    "type": "Edm.GeographyPoint"
  }

When the input data doesn't match the schema of the target index, you can use a field mapping to reshape your data during the indexing process.

"fieldMappings" : [
    { "sourceFieldName" : "source_geolocation", 
      "targetFieldName" : "my_index_geolocation" }
]

The indexer expects the representation of values to follow the GeoJSON point type format

"source_geolocation": {    
    "type": "Point",    
    "coordinates": [125.6, 10.1]   }

Recognize non-labeled location data in text

If the source data is unstructured and location information is available in the content of a document - compared to the previous case with a fixed schema - there are multiple ways to extract the information from documents or images. This mostly involves creating a custom skill to call in order to parse the text and recognize the geolocation data. It could be as simple as parsing/searching the content of documents for a certain keyword such as "Position", "Location", "lon", "lat" and transforming it to GeoJSON format. The custom skill can be added to the skillset to enrich the Azure Search pipepline.

A powerful tool one could use for translating address data into geolocation information is one of the Azure Search Power Skills: GeoPointFromName. It expects a key named "address" in the input dictionary containing the location information. Azure Search has a set of predefined skills, one of which is Entity Recognition. It extracts entities of different types from text and uses the machine learning models provided by Text Analytics in Cognitive Services. Assuming your address information has been extracted earlier in the pipeline i.e. through extracting the Entity "locations", it can be piped into the custom API by adding a custom skill. Here's an extract of the skillset containing Entity Recognition and the custom API call:

 # Extract entities
        {
          "@odata.type": "#Microsoft.Skills.Text.EntityRecognitionSkill",
          "categories": [ "Organization", "location", "person", "datetime", "url" ],
          "defaultLanguageCode": "en",
          "inputs": [
            {
              "name": "text",
              "source": "/document/content"
            }
          ],
          "outputs": [
            {
              "name": "locations"
            },
            {
              "name": "persons"
            },
            {
              "name": "urls"
            },
            {
              "name": "entities"
            }
          ]
        },
# Use a custom skill to transform the location into a geolocation point 
	 {
    	"@odata.type": "#Microsoft.Skills.Custom.WebApiSkill",
    	"description": "Geo point from name",
    	"context": "/document/content/locations/*",
    	"uri": "[AzureFunctionEndpointUrl]/api/geo-point-from-name?code=		[AzureFunctionDefaultHostKey]",
 	   "batchSize": 1,
    	"inputs": [
        	{
            "name": "address",
            "source": "/document/content/locations/*"
        	}
    	],
    	"outputs": [
        	{
            "name": "geolocation",
            "targetName": "geolocation"
        	}
    	],
    "httpHeaders": {}
	}

Add the output for the geolocation to "outputFieldMappings" and run the indexer.

    {
      "sourceFieldName" : "/document/geolocation", 
      "targetFieldName" : "my_index_geolocation"
    }