Implement information extraction solutions 5.1 Build retrieval and grounding pipelines

PDFs in Blob Storage contain scanned diagrams with printed and handwritten labels.

Study guide skill: Configure RAG ingestion flow, including documents and using optical character recognition (OCR)

Question B1-047

PDFs in Blob Storage contain scanned diagrams with printed and handwritten labels. The text in those images must be searchable together with the rest of each document. What should the enrichment pipeline do?

Select one answer.

  • A Enable semantic ranker on the content field
  • B Add a vectorizer to the index
  • C Set the blob indexer's imageAction to generate normalized images, run the OCR skill on /document/normalized_images/*, and merge the extracted text into the content ✓ Correct answer
  • D Set textExtractionAlgorithm to handwritten on the OCR skill and skip image extraction

Answer: C

The OCR skill extracts printed and handwritten text from image files and embedded images. It works on /document/normalized_images, which the blob indexer produces when imageAction isn't none, and a merge step combines the text with the document content. textExtractionAlgorithm is deprecated, and ranking or vectorizers don't extract text.

Deep dive

How it works

The OCR skill recognizes printed and handwritten text in image files. It extracts text from standalone images and from images embedded in PDFs or other files, using the Azure Vision Read API for supported languages. Its input only works with the /document/normalized_images field, which the blob indexer produces when imageAction is set to a value other than none.

The pipeline for the scenario looks like this:

  1. On the indexer, set imageAction so embedded images are extracted as normalized images.
  2. Run the OCR skill with the context /document/normalized_images/*. It outputs plain text, plus layoutText with locations.
  3. Use the Text Merge skill to merge the OCR text into the document's content. A common use is merging text from images into the content field when a PDF or Word document mixes text and embedded images.

The skill needs a billable Foundry resource beyond 20 documents per indexer per day, and image extraction is billed by Azure AI Search.

The other options, and when you'd use them

  • Enable semantic ranker on the content field: Semantic ranker is a set of query-side capabilities that rerank search results with language understanding models. That's a ranking step at query time, not text extraction.
  • Add a vectorizer to the index: A vectorizer handles query-time vectorization, turning text or images into vectors. That's a different job from extracting text out of images.
  • textExtractionAlgorithm set to handwritten, skip image extraction: That parameter is deprecated because the current Read API extracts printed and handwritten text at once. Without image extraction, OCR has no normalized images to read.

Exam tip

For text inside images, think "imageAction, then OCR on normalized_images, then merge". Ranking and vectorizer options are distractors at this stage.

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 Information extraction questions · AI-103 study guide

Thanks for your feedback!