AI Enrichment with Skillsets

Quick Answer: Skillsets define AI enrichment pipelines using built-in skills (OCR, NER, key phrases, sentiment) and custom skills (Web API calls). Skills read from and write to the enrichment tree (hierarchical document structure). Knowledge stores persist enriched data outside the search index.

Skillset Structure

{
    "name": "my-ai-skillset",
    "description": "Extract entities, key phrases, and perform OCR",
    "skills": [
        {
            "@odata.type": "#Microsoft.Skills.Vision.OcrSkill",
            "name": "ocr-skill",
            "context": "/document/normalized_images/*",
            "inputs": [
                {"name": "image", "source": "/document/normalized_images/*"}
            ],
            "outputs": [
                {"name": "text", "targetName": "ocrText"}
            ]
        },
        {
            "@odata.type": "#Microsoft.Skills.Text.V3.EntityRecognitionSkill",
            "name": "entity-recognition",
            "context": "/document",
            "categories": ["Organization", "Person", "Location"],
            "inputs": [
                {"name": "text", "source": "/document/content"}
            ],
            "outputs": [
                {"name": "organizations", "targetName": "organizations"},
                {"name": "persons", "targetName": "people"},
                {"name": "locations", "targetName": "locations"}
            ]
        },
        {
            "@odata.type": "#Microsoft.Skills.Text.KeyPhraseExtractionSkill",
            "name": "key-phrase-extraction",
            "context": "/document",
            "inputs": [
                {"name": "text", "source": "/document/content"}
            ],
            "outputs": [
                {"name": "keyPhrases", "targetName": "keyPhrases"}
            ]
        }
    ],
    "knowledgeStore": {
        "storageConnectionString": "<storage-connection-string>",
        "projections": [
            {
                "tables": [
                    {
                        "tableName": "documentsTable",
                        "generatedKeyName": "docId",
                        "source": "/document"
                    }
                ]
            }
        ]
    }
}

Built-in Skills

Text Skills

Vision Skills

Utility Skills

The Enrichment Tree

The enrichment tree is a hierarchical structure representing the document and its enrichments:

/document
├── content                        (original text content)
├── metadata_storage_name          (file name)
├── metadata_storage_path          (file path)
├── normalized_images/             (extracted images)
│   ├── [0]
│   │   ├── ocrText               (OCR output)
│   │   └── imageTags             (image analysis output)
│   └── [1]
│       ├── ocrText
│       └── imageTags
├── organizations                  (NER output)
├── people                         (NER output)
├── locations                      (NER output)
├── keyPhrases                     (key phrase output)
└── language                       (language detection output)

On the Exam: Understanding the enrichment tree path syntax is critical. Skills reference paths like /document/content (text content), /document/normalized_images/* (each image), and /document/organizations (enrichment output). Questions may ask you to fix a skillset by correcting the input/output paths.

Custom Skills (Web API Skill)

Custom skills call external HTTP endpoints for processing not covered by built-in skills:

{
    "@odata.type": "#Microsoft.Skills.Custom.WebApiSkill",
    "name": "custom-classification",
    "description": "Call a custom classification API",
    "uri": "https://my-function-app.azurewebsites.net/api/classify",
    "httpMethod": "POST",
    "timeout": "PT30S",
    "batchSize": 10,
    "context": "/document",
    "inputs": [
        {"name": "text", "source": "/document/content"}
    ],
    "outputs": [
        {"name": "category", "targetName": "documentCategory"}
    ]
}

Custom Skill Request/Response Contract

The custom endpoint must accept and return data in a specific format:

Request:

{
    "values": [
        {
            "recordId": "1",
            "data": {
                "text": "Document content to classify..."
            }
        }
    ]
}

Response:

{
    "values": [
        {
            "recordId": "1",
            "data": {
                "category": "Legal"
            },
            "errors": [],
            "warnings": []
        }
    ]
}

Knowledge Stores

Knowledge stores persist enriched data outside the search index for analytics and downstream processing:

Projection Types

Use Cases

• Power BI dashboards: Connect to table projections for business analytics
• Data science: Use object projections as training data for ML models
• Document archives: Store enriched documents with extracted metadata