Implement information extraction solutions 5.2 Extract content from documents
You ingest scanned PDFs into a search pipeline with Content Understanding.
Study guide skill: Produce clean, grounded representations to use with agents and RAG by using Content Understanding
Question B5-022
You ingest scanned PDFs into a search pipeline with Content Understanding. You need words, paragraphs, tables, sections, and figure locations, and the analyzer must not require any language model or embedding model deployment. Which prebuilt analyzer should you use?
Select one answer.
Answer: A
prebuilt-layout extracts words, figures, paragraphs, and tables, identifies document structure such as sections, and doesn't require a language model or embedding model. prebuilt-read provides basic OCR without layout analysis, and prebuilt-digitalParse reads a file's internal structure, so scanned documents should use read or layout instead. prebuilt-documentSearch is a RAG analyzer that adds figure descriptions and a one-paragraph document summary, and unlike layout, read, and digitalParse, the page doesn't describe it as working without a language model or embedding model.
Deep dive
How it works
Content Understanding's content extraction prebuilts differ in how much structure they return and whether they need a model deployment. The constraint here is "no language model or embedding model", so look at the prebuilts whose descriptions say they don't require one. Among those, the need for sections and figures picks the analyzer:
- prebuilt-layout: extracts words, figures, paragraphs and tables, and identifies document structure including sections and formatting. It doesn't require a language model or embedding model. That covers every item in the question.
- prebuilt-read: basic OCR. Words, paragraphs, formulas and barcodes, but no layout analysis, so no sections or figure locations.
- prebuilt-digitalParse: reads machine-readable content straight from the file's internal structure and encoding. For scanned or image-based documents, the page says to use read or layout.
Scanned PDFs + structure + no models = prebuilt-layout. If you later want figure descriptions, summaries or chunked output for vector indexing, prebuilt-documentSearch adds those on top of the layout content.
The other options, and when you'd use them
- prebuilt-documentSearch: It adds generated output such as detailed figure descriptions and a one-paragraph document summary. That goes beyond extraction; it's the RAG ingestion analyzer, not the model-free one.
- prebuilt-digitalParse: It analyzes the file's internal encoding. For scanned documents, use prebuilt-read or prebuilt-layout.
- prebuilt-read: It provides foundational text extraction without layout analysis, so no sections or figures.
Exam tip
Content Understanding without models: read = text only, layout = text + structure + figures, digitalParse = born-digital files only.
References
Microsoft Learn references
Last reviewed: