Extract Data from Documents with Azure AI

Azure Document Intelligence REST SDK for TypeScript is a skill for extracting text, tables, and structured data from documents using Azure's prebuilt and custom models. It is a REST client, not a class-based SDK - DocumentIntelligence is a function that returns a client, authenticated via DefaultAzureCredential or an API key. Once created, the client analyzes documents (by hosted URL or a local file read and base64-encoded) with any of nine prebuilt models (read/OCR, layout, invoice, receipt, ID document, W-2 tax form, health insurance card, contract, bank statement), extracts named fields like invoice vendor and total, lists models via a paginate() helper, builds custom models in template or neural mode, builds and runs document classifiers with per-type training data, and queries service info such as the custom-model limit and count via the /info endpoint.

Use it when a TypeScript or Node project needs to pull structured fields - vendor name, total, due date, line items - out of documents like invoices, receipts, ID cards, tax forms, or contracts, either with Azure's prebuilt models or a custom-trained one. It is specifically a REST client on @azure-rest/ai-document-intelligence; reach for it when that is the SDK in use rather than assuming the older class-based pattern, since DocumentIntelligence here is a function, not a class.

Input is a document, given either as a urlSource (a hosted URL) or a base64Source (a local file read and base64-encoded), an Azure DOCUMENT_INTELLIGENCE_ENDPOINT, and either DefaultAzureCredential or a DOCUMENT_INTELLIGENCE_API_KEY. Output is an AnalyzeOperationOutput reached by polling a long-running operation to completion, exposing pages, tables, and, for prebuilt models, named fields with confidence scores, for example invoice.fields?.VendorName?.content or receipt.fields?.Total?.content - the skill's own guidance is to set thresholds on those confidence scores rather than trust every extracted field blindly.

npm install @azure-rest/ai-document-intelligence @azure/identity

import DocumentIntelligence from "@azure-rest/ai-document-intelligence";
import { DefaultAzureCredential } from "@azure/identity";

const client = DocumentIntelligence(
  process.env.DOCUMENT_INTELLIGENCE_ENDPOINT!,
  new DefaultAzureCredential()
);

Prebuilt models cover OCR (prebuilt-read), layout with tables and selection marks (prebuilt-layout), invoices, receipts, ID documents, W-2 tax forms, health insurance cards, contracts, and bank statements. Custom models can be built in template or neural mode from blob-stored training data - the skill recommends neural mode since it handles more variation than template. Custom classifiers can split and route documents by doc type using an azureBlobSource per type. The skill's own best practices: always drive analysis through getLongRunningPoller() since it is async, check isUnexpected() as a type guard on every response, prefer prebuilt models when they fit and reserve custom models for specialized documents, and use the paginate() helper when listing models.

TypeScript and Node developers building document-processing pipelines on Azure. Most reach for the high-level poller helpers above; developers writing lower-level polling logic of their own can also use parseResultIdFromResponse and the other exported types (DocumentClassifierBuildOperationDetailsOutput) directly from the package.