AI-901 lesson 11: Content Understanding
Azure Content Understanding in Foundry Tools turns unstructured content, meaning documents, images, audio and video, into structured output that an application can use. The unit of work is the analyzer: a saved definition of how one kind of content is processed and which fields to return. Microsoft ships prebuilt analyzers for common cases, and a custom analyzer can be defined when the fields are specific to a business. The service is part of the Foundry resource, so it fits next to the other tools in a project.
What an analyzer does
The figure above shows the flow. Content goes into an analyzer; the analyzer applies its settings; structured results come out to the application. Inside, the work has two layers. Content extraction turns the raw input into usable text and structure: optical character recognition for scanned pages, layout such as paragraphs and tables, and transcription for speech in audio and video. Field extraction then produces named values according to a schema, such as an invoice number or a call topic.
Fields can be filled in three ways in the documentation: extracting a value as it appears in the source (documented for documents), classifying content into categories, or generating a value such as a summary. Results can include confidence scores and a pointer to where in the source a value came from, which supports review. Output may be delivered as Markdown for search and retrieval or as structured data matching the schema.
Each input type
| Input | Typical result |
|---|---|
| Documents and forms | Text, layout, tables and named fields such as totals, dates and parties |
| Images | A description and insights about visual content, or fields defined for a custom case |
| Audio | A transcript, a summary and fields such as topic or sentiment of a conversation |
| Video | Transcript and a description of each segment, with scene or topic splits |
Prebuilt analyzers exist for several kinds of document, including invoices, receipts, identity documents and contracts, and for call-centre recordings. Others are aimed at search and retrieval over documents, images, audio and video. If no prebuilt analyzer matches, a custom analyzer starts from a base analyzer for the modality and adds the fields wanted. Documentation notes that features marked preview in newer API versions can change, so treat preview-only capabilities with care.
Worked example: a 200-person logistics company
The company receives supplier invoices as PDFs and photographed delivery notes, and records customer calls. Finance wants invoice number, date and total in the accounting system. Operations wants the delivery note’s reference number. Support wants each call’s topic and sentiment.
The invoices have a prebuilt analyzer, so the app uses it unchanged. Delivery notes follow an in-house layout, so a custom analyzer is defined from the document base with a field for the reference number. Calls use an audio analyzer returning a transcript, summary and the requested fields.
The lightweight application does the same thing each time: connect to the Foundry resource, pick the analyzer, submit the file, wait for the analysis to finish, then read the fields and confidence values and write them to a database. Low-confidence values go to a person.
When extraction beats a chat model
A chat model can read an invoice if asked, but extraction is stronger when the same fields are needed every time, at volume, with traceable sources and scores. A chat model suits one-off questions about content. A useful test: if the output goes into a database column, think extraction; if it goes to a reader as prose, think generation.
Easy to confuse
- Content Understanding vs speech recognition: speech recognition only transcribes; Content Understanding can also return structured fields from the audio.
- Content Understanding vs image interpretation by a model: one returns a defined schema, the other free-form text.
- Prebuilt vs custom analyzer: use prebuilt when the document type is covered, custom when your fields are unique.
Key points
- An analyzer defines how content is processed and what fields come back.
- One service handles documents, images, audio and video.
- Prebuilt analyzers cover common documents and recordings; custom analyzers add business-specific fields.
- Results are structured fields, optionally with confidence scores and source locations.
- A lightweight app submits content to an analyzer, waits for the result and reads the fields.
- Prefer extraction over chat when the same fields are needed repeatedly.
Check yourself
Question 1. A 90-person estate agency scans rental contracts and needs tenant name, start date and monthly rent as fields for its database. Which tool fits best?
- A. Speech synthesis
- B. An image-generation model
- C. A Content Understanding analyzer
- D. Sentiment analysis
Show answer
Answer: C
Pulling named fields from documents is the core use of Content Understanding. Speech synthesis produces audio, image generation creates pictures, and sentiment analysis rates opinions in text rather than returning contract fields.
Want more questions like this? Full AI-901 practice tests →
Question 2. A training firm stores recorded workshops and wants each one segmented, with a transcript and a short description of every segment. Which approach matches?
- A. Image generation
- B. Content Understanding with a video analyzer
- C. Entity detection on the file
- D. A text-only prompt containing the file name
Show answer
Answer: B
Content Understanding handles video, returning transcripts and per-segment descriptions. Image generation makes new pictures, entity detection works on text only, and a text-only prompt cannot process a video file.
Want more questions like this? Full AI-901 practice tests →
Question 3. A team's supplier forms follow an unusual in-house layout with a field no prebuilt analyzer returns. What should the team do?
- A. Rely on a prebuilt invoice analyzer unchanged
- B. Use speech recognition
- C. Use an image-generation model
- D. Define a custom analyzer based on the document base
Show answer
Answer: D
A custom analyzer lets the team start from the document base and define the missing field. A prebuilt analyzer will not return fields it was not designed for, speech recognition is for audio, and image generation would not read the forms at all.
Want more questions like this? Full AI-901 practice tests →
This lesson covers the information-extraction part of the Implement AI solutions by using Microsoft Foundry domain (55-60% of AI-901). The Foundry implementation recap shows how it fits with the other solution types, and RAG explained shows why extracted Markdown is useful for search.