Implement information extraction solutions 5.2 Extract content from documents

A Content Understanding invoice analyzer returns low-confidence values for the delivery…

Study guide skill: Implement analyzers for generating structured or markdown outputs for downstream reasoning by using Content Understanding

Question B5-021

A Content Understanding invoice analyzer returns low-confidence values for the delivery date field on invoices from several suppliers, and each supplier labels that date differently. Following Microsoft's best practices, what should you do first?

Select one answer.

A Add one labeled example invoice to the knowledge base, because a single example covers all templates
B Describe what the field isn't, for example 'not the invoice date and not the due date'
C Create a separate document category for each supplier's invoice layout
D Rewrite the field description with location hints, the expected formats, and every label the suppliers use, and add labeled examples for each template only if confidence stays low ✓ Correct answer

Answer: D

The guidance is to refine field descriptions before adding labeled training examples: include location hints, format expectations, and all the alternative labels the documents use. If confidence is still low, add training examples that cover each distinct layout or template, not just one example overall. Descriptions should say what a field is rather than what it isn't, and categories should be semantic, not one per visual template.

Deep dive

How it works

Content Understanding's best practices give a clear order of operations for weak fields: start with descriptions, then add examples. Prioritize refining field descriptions before adding labeled training examples, because clear descriptions often resolve issues without more data.

A strong description for the delivery date has three parts, all named in the guidance:

Location hints: where the date usually appears, such as the delivery information section.
Format expectations: for example MM/DD/YYYY or DD-MM-YYYY.
Alternative labels: list all possible names for the field, especially when you handle diverse templates, so the model recognizes it regardless of labeling.

Write it in affirmative language. Describe what the field is, not what it isn't.

If confidence is still low after that, add training examples to the knowledge base. Include an example for each distinct layout or template the analyzer must handle, not just one example overall, and test on a representative set before and after training so you can confirm the examples actually improved results.

The other options, and when you'd use them

One labeled example for all templates: The guidance says to include an example for each distinct layout or template, not just one example overall. And examples come after descriptions.
Describe what the field isn't: The best practices say to use affirmative language. Instead of "This field isn't the invoice date and isn't the due date," say what it is and where it's found.
A separate category per supplier layout: Use semantic categories, not visual templates. Don't create separate categories for the same document type with different layouts; use one Invoice category for all invoice variations.

Exam tip

Low confidence on a field = rewrite the description first (location, format, every label, affirmative wording), then add one example per template.

References

Best practices for using Content Understanding - Foundry Tools

Last reviewed: 10 Oct 2026

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →