AI-103 Practice Question 70: Document Intelligence Composed Model
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 7 · question 10 of 10
A logistics company built three Document Intelligence custom models, one each for its purchase orders, its bills of lading, and its customs declarations, each a distinct, consistent layout. Incoming files arrive at a single ingestion endpoint with no indication of which of the three document types each one is, and the volume is growing toward a fourth and fifth document type next quarter. The team wants one endpoint that automatically applies the right model to each document.
What should they build?
Show answer and explanation
Answer:
This is the scenario composed models exist for: "compose several custom models into a single model for easier data extraction from groups of distinct form types... your composed model might include custom models trained to analyze your supply, equipment, and furniture purchase orders. Instead of manually trying to select the appropriate model, you can use a composed model to determine the appropriate custom model for each analysis." A single call to the composed model ID classifies the incoming document and routes it to the matching component model, and the response's docType property reports which one was used, with room to assign up to 500 component models, comfortably covering the two new document types coming next quarter.
- A is wrong because training one neural model on mixed, structurally different layouts throws away the fact that each document type already has its own well-performing model; Microsoft's guidance instead directs training one model per document type and composing the results, not merging the training data. - B is wrong because it hand-rolls, in application code, the classification step a composed model already performs as a first-class, trainable capability, brittle keyword matching instead of a maintained classifier, with no docType output and no clean way to add new document types. - D is wrong on cost and design: it calls all three (soon five) models on every document and reconciles the results afterward, multiplying extraction cost by the number of models instead of routing each document to the one model that applies.