Information extraction

What this area tests

• extracting information from documents and forms
• extracting information from images
• extracting information from audio and video
• building a lightweight app with information extraction

All four use Azure Content Understanding in Foundry Tools.

What Content Understanding does

Content Understanding uses generative AI to process documents, images, video and audio into an output format you define. Instead of one service per file type, you use one service and choose an analyzer for the job.

Analyzers

An analyzer is the recipe for processing a file:

• Prebuilt analyzers are ready to use for common tasks. Some focus on content extraction (reading text and layout). Others are domain-specific, such as ID documents or invoices, and come with a predefined set of fields.
• Custom analyzers start from a base analyzer (document, image, audio or video) and add your own field schema: the list of fields you want back, with names, types and descriptions.

Microsoft’s advice for production: copy a prebuilt analyzer into a custom one rather than relying on the prebuilt version directly, because prebuilt definitions can change between API versions.

Two kinds of output

• Content extraction: the text and layout of a document, or a transcript of audio and video, often as markdown. This is ideal input for search and RAG.
• Field extraction: specific values defined by the schema, such as invoice number, total and due date, returned as structured data your app can use directly.

Exam patterns

• One service across many file types: Content Understanding.
• Specific values from forms or IDs: field extraction with a domain or custom analyzer.
• Searchable text from a large document collection: content extraction, then index it for search.
• Transcribing audio and pulling out structured details such as names and topics: Content Understanding with an audio analyzer, not speech to text alone.