AI Apps
Azure AI Document Intelligence
Extraction that reads scans, tables, and forms into structured text, which is what makes a document estate retrievable in the first place.
Product
Azure AI Document IntelligenceAzure AI Document Intelligence reads documents into structured output: text, key value pairs, tables, and layout, from digital files and from scans. It matters upstream of everything else, because a retrieval index built from a bad extraction returns bad passages no matter how good the model reading them is. Prebuilt models cover common documents such as invoices, receipts, and identity documents, and custom models are trained where the document is specific to the organisation. Layout analysis preserves the structure a table carries, so a number stays attached to the row and column that give it meaning.
How it works
- Documents arrive from storage, an upload, or a line of business system, in whatever mix of digital files and scans the estate actually holds.
- Optical character recognition reads scanned and photographed pages, including handwriting where the model supports it.
- Layout analysis records headings, reading order, and table structure, so a figure keeps the row and column that give it meaning.
- Prebuilt models extract the fields common documents carry, and custom models are trained where a document is specific to the organisation.
- The structured output is split into passages sized for retrieval and passed to indexing, carrying the source location for citation.
- Confidence scores on each extracted field mark what a human should check rather than leaving the workload to assume it was right.
- Low confidence extractions are routed for review, and the corrections become training data for the custom model.
A problem it solves
The knowledge an assistant needs is in ten years of scanned contracts and forms, so a retrieval index built from a plain text dump loses the tables and returns passages that answer nothing. Azure AI Document Intelligence reads those documents into structured output with their layout intact, so what reaches the index is a passage that still makes sense on its own.