Implement computer vision solutions 3.2 Design and implement multimodal understanding workflows

You create a custom Content Understanding analyzer with baseAnalyzerId set to…

Study guide skill: Configure single-task and pro-mode Content Understanding pipelines

Question B5-008

You create a custom Content Understanding analyzer with baseAnalyzerId set to prebuilt-image for production-line photos. It must return DefectType (scratch, crack, or misalignment) and a short DefectSummary. A developer defined DefectType with method extract and estimateSourceAndConfidence set to true so that low-confidence results could go to human review. Which change is correct?

Select one answer.

  • A Keep method extract and also set enableOcr to true in the analyzer configuration
  • B Change baseAnalyzerId to prebuilt-video, which supports extract for still images
  • C Use method classify with an enum of the three values for DefectType and method generate for DefectSummary, and don't rely on confidence scores, which only document analyzers support ✓ Correct answer
  • D Use method generate for DefectType, because generate returns a confidence score for each value

Answer: C

In Content Understanding, the extract method is only supported for document analyzers, and so are confidence scores, so an image analyzer should use classify for a fixed set of values and generate for values that need interpretation. Image analyzers support only the returnDetails configuration option, while enableOcr is a document analyzer option. Each base analyzer matches a content type, so prebuilt-image is the right parent for photos.

Deep dive

How it works

Every field in a Content Understanding schema has a method, and the best-practices guide says to set it explicitly by purpose:

  • Generate: values that need inference or summarization, such as a risk level or a summary.
  • Classify: a selection from predefined options, best used with an enum of fixed values.
  • Extract: values that appear directly in the content, such as an invoice number. Extract is only supported for document analyzers.

For production-line photos on prebuilt-image, DefectType is a fixed choice (scratch, crack, misalignment), so it's classify with an enum. DefectSummary needs interpretation, so it's generate.

Confidence works the same way. estimateSourceAndConfidence, which returns source location and a confidence score per field, is supported by document analyzers. An image analyzer can't feed a confidence-threshold review queue this way. Image analyzers support only one configuration option, returnDetails. If you need human review, base it on something else, such as sampling results or reviewing every "crack" result.

prebuilt-image remains the right parent. Base analyzers provide core processing for each content type, and photos are images.

The other options, and when you'd use them

  • Keep extract and set enableOcr: Extract is only for document analyzers. enableOcr is a document analyzer option; image analyzers support only returnDetails.
  • Change base to prebuilt-video: Base analyzers match content types, and extract is still document-only. A video base doesn't make still photos extractable.
  • generate returns a confidence score: Source and confidence estimation is listed as supported by document analyzers, not image analyzers.

Exam tip

Content Understanding image analyzer: classify (enum) or generate, never extract, and no confidence scores. Extract and confidence = document analyzers.

References

Microsoft Learn references

Last reviewed:

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

All Computer vision questions · AI-103 study guide

Thanks for your feedback!