Turning a scanned contract into a JSON object is not just optical character recognition. Documents can contain tables, checkboxes, handwritten annotations, footnotes, amendments, signatures, and multiple values that share the same label. A useful extraction pipeline must identify which field matters, where it came from, how confident the system is, and when it should refuse to decide. AI-103 assesses information extraction as a distinct domain because downstream applications and agents depend on these details.<\/p>\n
Consider a supplier onboarding system. Each supplier uploads an invoice, a certificate, and a short contract. The application must identify supplier name, invoice date, purchase-order number, total before tax, applicable tax, insurance expiry, and an exception clause. A PDF may contain two totals; a photographed form may have a handwritten correction. A plausible field value without page evidence is not enough to approve payment.<\/p>\n
Azure Content Understanding<\/a> in Foundry Tools can process documents, images, audio, and video into structured representations. It complements rather than automatically replaces deterministic OCR and Document Intelligence extraction. Choose a pipeline based on source types, required fields, latency, cost, accuracy, and the review consequences of failure.<\/p>\n List accepted file types, maximum size, privacy classification, expected languages, encryption and retention needs, and whether documents are single files or collections. Assign stable identifiers at ingestion and retain the source revision. Never rely on the uploaded filename as proof of origin or authority. Validate media type and content independently where possible and reject unsupported or suspicious inputs before processing.<\/p>\n Define a field schema with names, types, requirements, permitted nulls, and provenance. For an invoice, purchase-order number is a string rather than a number, because leading zeros and letters matter. A date should be represented in a consistent format with ambiguity handling for different regional conventions. A total should carry currency and perhaps a source cell or span reference. If two totals are present, choose only when the field definition clearly disambiguates them.<\/p>\n Specify review rules in the contract. An uncertain invoice total above a threshold may require a human check, while a missing optional phone extension might not. A model-produced confidence score is not a substitute for a calibrated decision threshold. Build acceptance examples and counterexamples that demonstrate what “needs review” looks like before you start adjusting prompts or analyzers.<\/p>\n OCR identifies text characters and their positions. Layout analysis identifies structures such as paragraphs, tables, headers, and reading order. Field extraction maps this structure to a business schema. Semantic interpretation may reason across clauses or reconcile a form field with a related document. These stages solve different problems and should be evaluated separately.<\/p>\n A two-column contract may contain a term in the left column and its exception in the right. OCR can recognize both strings while a naive reading order places the exception under the wrong clause. Likewise, a table may contain subtotals and a final total; extracting a number accurately does not prove it represents the requested financial field. Preserve layout boundaries and table relationships when constructing normalized text.<\/p>\n Microsoft’s document extraction<\/a> guidance describes multimodal analyzers and the relationship to Document Intelligence. Where deterministic extraction is reliable and sufficient, it may be preferable to a more general generative process. Where documents vary widely in format or require semantic fields, a customized Content Understanding analyzer may be justified. Compare outcomes and failure modes on a representative dataset.<\/p>\n As of Microsoft’s September 2026 description, Content Understanding API version 2025-11-01 is generally available for core document, image, audio, and video extraction. Version 2026-06-01-preview offers additional capabilities, including agentic document reasoning and richer classification or segmentation behaviors. These are different operational commitments, not just alternate endpoints with identical guarantees.<\/p>\n An analyst may want an agentic extractor to reason over a difficult table or perform a validation calculation. That can be useful in a lab, but the architecture must label the preview status and compare with simpler supported approaches. Do not present a feature as universally GA merely because the portal has a button for it. Check region availability, model prerequisites, quotas, and the service’s current version notes.<\/p>\n The analyzer reference<\/a> explains base, RAG-oriented, domain-specific, and custom analyzers, plus output options. Configuration determines whether text, tables, images, segments, or user-defined fields are produced. Pin the analyzer ID and version with your application tests and avoid depending on fields that are absent under the chosen API.<\/p>\n The April 2026 AI-103 objective list still refers to \u201csingle-task and pro-mode\u201d Content Understanding pipelines, but production terminology has changed since that guide was written. Microsoft’s former pro mode belonged to the retired If an invoice’s total is returned as 24,500, the application should know which document, page, table, row, and field supports the value. For a contract clause, preserve the paragraph and relevant exception. Source locations let reviewers verify facts and let regression tests distinguish an extraction defect from a downstream summarization defect. Keep references to original documents where retention and permissions allow.<\/p>\nDesign an input and field contract before selecting an analyzer<\/h2>\n
Distinguish OCR, layout, fields, and semantic interpretation<\/h2>\n
Choose GA and preview features deliberately<\/h2>\n
2025-05-01-preview<\/code> API; it is not a current GA deployment option. For straightforward extraction, prefer the supported 2025-11-01<\/code> API. The newer agentic mode<\/a> uses 2026-06-01-preview<\/code> and is limited to document analyzers and a single input file per request in its initial release. It is not a drop-in replacement for legacy multi-document pro mode: validate workflow configuration, field methods, privacy, latency, and result shape before migration. This is precisely why exam study notes and working integration code must be versioned separately.<\/p>\nPreserve traceable evidence for agents and search<\/h2>\n