Information Extraction — Azure AI Apps and Agents Developer AI-103 practice questions

How do the layout model and the prebuilt invoice model in Azure Document Intelligence in Foundry Tools differ?

1. Layout returns labeled invoice fields such as vendor name and total due
2. They return identical results; only the model ID used in the call differs
3. Layout returns text, tables and structure; prebuilt invoice returns fields
4. The prebuilt invoice model returns raw text only, with no field extraction

Answer: C — Layout returns text, tables and structure; prebuilt invoice returns fields

C) Right — layout is a general structure extractor, while the prebuilt invoice model emits typed, labeled invoice fields with confidence scores. A) Reverses the two: layout has no invoice schema. B) The models have genuinely different output shapes. D) The prebuilt invoice model returns fields in addition to text.

Which statement about the Read OCR capability and PDFs is true?

1. It processes only the first page of any submitted PDF file
2. It requires a separately trained classification model first
3. It accepts PNG images only, so every PDF must be converted first
4. It accepts multi-page PDFs and TIFFs with a page range filter

Answer: D — It accepts multi-page PDFs and TIFFs with a page range filter

D) Multi-page PDF and TIFF input is supported and you can restrict processing to a page range. A) Multi-page documents are handled in one operation. B) No training step is involved in prebuilt OCR. C) Several image formats and PDFs are accepted.

A team is deciding between a Content Understanding analyzer and a custom Document Intelligence extraction model for a new workload. Which consideration MOST favors the analyzer?

1. The output must include per-field bounding regions on a fixed form template
2. The workload is a single fixed paper form processed at very high volume
3. The team already maintains a large labeled corpus for that exact form layout
4. The inputs span PDFs, images and audio recordings within the same workload

Answer: D — The inputs span PDFs, images and audio recordings within the same workload

D) Correct — mixed modalities in one workload is the differentiator, since Content Understanding spans documents, images, audio and video while Document Intelligence is document-oriented. A) points to fixed-form field localization, again squarely Document Intelligence territory. B) describes the classic custom extraction case. C) means the labeling investment already exists, which favors the custom model.

In Azure Content Understanding in Foundry Tools, what does defining an analyzer primarily accomplish?

1. It provisions a dedicated compute cluster sized for the expected document volume
2. It declares the output schema and fields to extract from a given content type
3. It fine-tunes an underlying model on the customer's uploaded content corpus
4. It configures a network route between the storage account and the service endpoint

Answer: B — It declares the output schema and fields to extract from a given content type

B) Correct — an analyzer is the configuration object that says what content type you are processing and what fields and shape the result should take. A) confuses a declarative configuration with infrastructure provisioning; capacity is handled by the service. C) tempts because field definitions feel like supervision, but analyzers are schema-driven rather than a training job. D) describes networking setup, which is a separate resource concern.

Ten thousand purchase receipts per day must yield merchant, date, total and tax as typed values written to a relational table, with per-field confidence for audit. Which choice is MOST appropriate?

1. A multimodal prompt asking a generative model to read each receipt and return JSON
2. An analyzer producing markdown, parsed downstream into the relational columns
3. The prebuilt receipt model of Azure Document Intelligence in Foundry Tools
4. The layout model, mapping detected key-value pairs onto the relational columns

Answer: C — The prebuilt receipt model of Azure Document Intelligence in Foundry Tools

C) Correct — a high-volume, fixed-schema extraction with a required per-field confidence score is the case Document Intelligence prebuilt models serve best, at the lowest cost per page. A) tempts because a multimodal model can read a receipt, but it gives no calibrated per-field confidence and costs far more at this volume. B) produces prose-like structure that then needs fragile parsing into typed columns. D) returns generic structure rather than receipt field semantics, leaving the mapping to you.

Which service best extracts printed and handwritten text from a scanned PDF?

1. The Read OCR capability of Azure Vision in Foundry Tools
2. An image classification model trained in the Foundry portal
3. The Face capability of Azure Vision in Foundry Tools
4. Azure Translator in Foundry Tools with a document glossary

Answer: A — The Read OCR capability of Azure Vision in Foundry Tools

A) Read handles printed and handwritten text across multi-page documents and returns the text with coordinates. B) Classification labels an image, it does not transcribe it. C) Face works on faces, not text. D) Translation acts on text that has already been extracted.

A travel operator must read passport and driving licence images to capture holder name, document number and expiry date. Which prebuilt model of Azure Document Intelligence in Foundry Tools is MOST appropriate?

1. The receipt model, which is tuned for small printed documents with dense fields
2. The layout model, which returns text, tables, and selection marks with coordinates
3. The read model, which returns text and reading order across multiple languages
4. The identity document model, which returns holder and document fields directly

Answer: D — The identity document model, which returns holder and document fields directly

D) Correct — the identity document model is purpose-built for passports and licences and returns exactly those typed fields. A) tempts by document size but is trained on purchase receipts and would emit merchant and total fields. B) gives structure without identity field semantics. C) gives text without any field mapping, leaving all the interpretation to you.

A vendor supplies purchase orders in a layout that a prebuilt model covers only partially: most fields resolve, but three vendor-specific codes are never returned. Which approach is the MOST appropriate?

1. Abandon the prebuilt model and label every field from scratch in a custom project
2. Raise the confidence threshold until the vendor-specific codes begin to appear
3. Keep the prebuilt call and add a custom extraction model for the vendor-specific codes
4. Ask the vendor to redesign the purchase order to match the prebuilt model's schema

Answer: C — Keep the prebuilt call and add a custom extraction model for the vendor-specific codes

C) Correct — combining a prebuilt call for the common fields with a targeted custom model for the fields it does not know keeps labeling effort proportional to the gap. A) tempts as the clean single-model answer, but relabeling fields the prebuilt model already handles wastes effort. B) misunderstands thresholds: a field the model was never trained to emit will not appear at any threshold. D) shifts the problem onto a party you usually cannot direct.

A pipeline extracts values that are later used to approve payments. Which practice BEST balances throughput against error cost?

1. Review every document by hand regardless of score, so that correctness is guaranteed
2. Set one global confidence threshold and apply it identically to every extracted field
3. Set field-specific thresholds, with stricter gates on high-impact fields such as amounts
4. Trust all extractions and sample one percent of documents after payment has run

Answer: C — Set field-specific thresholds, with stricter gates on high-impact fields such as amounts

C) Correct — error cost is not uniform across fields, so tightening the gate where a mistake is expensive and loosening it elsewhere gives the best throughput for a given risk. A) is safe but discards automation entirely. B) tempts as the simple option, yet a threshold strict enough for payment amounts will needlessly route low-risk fields to review. D) detects errors only after money has moved, which is the wrong side of the control.

Which situation LEAST justifies choosing a pro-mode pipeline over a single-task analyzer?

1. Extracting five fixed fields from one standalone document per request
2. Cross-checking a submitted form against an attached supporting document
3. Grounding an extraction against a supplied reference data set
4. Producing one consolidated result from several related input files

Answer: A — Extracting five fixed fields from one standalone document per request

A) Correct — a single input with a fixed field list is the canonical single-task case, and pro mode would add cost and latency for no benefit. B) needs comparison across two inputs, which is a pro-mode strength. C) is exactly what reference data support in pro mode is for. D) requires consolidating multiple inputs into one result, again a pro-mode scenario.

Optical character recognition vs Azure Document Intelligence in Foundry Tools?

OCR returns the raw text and its position from an image or PDF. • Document Intelligence returns structured results: named fields, key-value pairs, tables and selection marks, with confidence scores.

What does a field-level confidence score mean, and how should it be used?

It expresses the model's certainty about that specific extracted value, not the document's overall correctness. Set a per-field threshold: auto-process above it, route below it to human review. Thresholds should be tuned per field, since a total needs more certainty than a description.

What is an analyzer in Azure Content Understanding in Foundry Tools?

An analyzer is a reusable, schema-driven extraction configuration: you define the fields you want and their types, and the analyzer applies that schema to incoming content, returning the populated fields with confidence. The same analyzer definition can be versioned and applied across a whole content set.

What distinguishes pro mode from a single-task analyzer in Content Understanding?

A single-task analyzer extracts from one input in one pass. Pro mode runs a multi-step reasoning pipeline that can consider several inputs together and reference supplied context or reference data, so it can reconcile and cross-check across documents. It costs more and takes longer, so it is reserved for cases where cross-document reasoning is the requirement.