Extract Data from Documents with Azure AI

This skill covers the Azure Document Intelligence REST SDK for TypeScript (@azure-rest/ai-document-intelligence), used to extract text, tables, and structured data from documents with prebuilt or custom models. DocumentIntelligence is a function, not a class, returned by the package and authenticated with either DefaultAzureCredential or an API key. Analysis is always async: post to a path like /documentModels/{modelId}:analyze with a urlSource or a base64-encoded local file, check the response with isUnexpected(), then wrap it in a poller from getLongRunningPoller() and call pollUntilDone() to get the final AnalyzeOperationOutput. Nine prebuilt models are listed: prebuilt-read (OCR), prebuilt-layout (text/tables/selection marks/structure), prebuilt-invoice, prebuilt-receipt, prebuilt-idDocument, prebuilt-tax.us.w2, prebuilt-healthInsuranceCard.us, prebuilt-contract, and prebuilt-bankStatement.us.

Use it when extracting structured fields from documents in a TypeScript/Node app - invoices, receipts, ID documents, tax forms, or your own document types via a custom model or classifier - and when you're comfortable with an async, poll-until-done REST workflow rather than a synchronous call. Prebuilt models cover common document types out of the box; reach for a custom model (built in template or neural mode from a labeled blob-storage training set) only when your documents don't match a prebuilt schema. A poller also exposes onProgress() if you want to log intermediate status while waiting for a long-running analyze, build, or classify operation to finish.

Configuration is DOCUMENT_INTELLIGENCE_ENDPOINT and DOCUMENT_INTELLIGENCE_API_KEY. A URL-based analyze call looks like this:

const initialResponse = await client
  .path("/documentModels/{modelId}:analyze", "prebuilt-layout")
  .post({
    contentType: "application/json",
    body: {
      urlSource: "https://example.com/document.pdf"
    },
    queryParameters: { locale: "en-US" }
  });

The resolved result exposes analyzeResult.pages, analyzeResult.tables, and for prebuilt-invoice/prebuilt-receipt a documents[0].fields object with named fields such as VendorName, InvoiceTotal, DueDate, MerchantName, Total, and a repeated Items collection with per-item Description/TotalPrice. Custom models and classifiers are built by posting to /documentModels:build or /documentClassifiers:build with an azureBlobSource pointing at labeled training data, then polled the same way; /documentModels lists existing models via the paginate() helper, and /info returns the account's custom-model limit and current count.

Authenticates via @azure/identity's DefaultAzureCredential or a plain API key, and is distributed as the npm package @azure-rest/ai-document-intelligence.

TypeScript/Node developers extracting document data who want prebuilt coverage for common document types plus a path to custom models or classifiers for their own - and who follow the SDK's own guidance to always poll with getLongRunningPoller, check isUnexpected for errors, apply confidence-score thresholds to extracted fields, and prefer neural mode over template mode for custom models that need to handle more layout variation.