Implement information extraction solutions 5.1 Build retrieval and grounding pipelines
An Azure AI Search skillset uses the Document Layout skill and the Text Split skill to…
Study guide skill: Implement enrichment by using custom or built-in skills for text, images, and layout
Question B5-027
An Azure AI Search skillset uses the Document Layout skill and the Text Split skill to chunk financial reports for RAG. Tables come out as plain text, and tables that span two pages are split apart, so answers miss figures. What should you use for a new skillset?
Select one answer.
Answer: A
The Azure Content Understanding skill outputs tables and figures in Markdown, recognizes cross-page tables as a single unit, and does both content extraction and chunking, while the Document Layout skill outputs tables as plain text, which can lose information. Microsoft recommends it for new skillsets, and the Document Layout skill remains supported for existing pipelines. Text overlap only preserves continuity between chunks, and the multimodal search page's skill comparison lists table extraction and preservation as not supported by the Document Extraction skill (and supported, including cross-page tables, by the Content Understanding skill).
Deep dive
How it works
For RAG over table-heavy documents, Azure AI Search's Azure Content Understanding skill (preview) fixes both problems in this pipeline:
- Table format: it outputs tables and figures in Markdown, which LLMs understand more easily. The Document Layout skill outputs them as plain text, which can lose information, such as which number belongs to which column.
- Tables across pages: it recognizes and extracts cross-page tables as a single unit, and chunks can span pages through semantic units, so a table isn't cut at the page break.
- Chunking: it does both content extraction and chunking, so you don't need the Text Split skill.
It uses the same interface as the Document Layout skill, so switching is mostly a skill swap. The multimodal search guidance is direct about which to pick: the Document Layout skill remains supported for existing pipelines, but for new skillsets, use the Azure Content Understanding skill.
One billing note: unlike the Document Layout skill, it doesn't include 20 free documents per indexer per day. It's charged at the Content Understanding price.
The other options, and when you'd use them
- Document Layout is recommended for new skillsets: The guidance says the reverse. Document Layout remains supported for existing pipelines; new skillsets should use the Content Understanding skill.
- Increase pageOverlapLength: Overlap only makes each page start with characters from the end of the previous one. Tables still come out as plain text and stay split.
- Document Extraction + Text Split: Document Extraction lets you control how built-in content extraction runs and names its output. It doesn't give Markdown tables or cross-page table extraction, which are the Content Understanding skill's advantages.
Exam tip
New RAG skillset with complex tables = Azure Content Understanding skill: Markdown tables, cross-page tables as one unit, built-in chunking, no Text Split.
References
Last reviewed: