Implement information extraction solutions 5.1 Build retrieval and grounding pipelines
Your RAG index uses the Text Split skill in pages mode.
Study guide skill: Configure RAG ingestion flow, including documents and using optical character recognition (OCR)
Question B3-030
Your RAG index uses the Text Split skill in pages mode. Answers often lack context because a key sentence is split across two chunks. Which setting addresses this?
Select one answer.
Answer: B
pageOverlapLength makes each page start with a number of characters or tokens from the end of the previous page, which preserves continuity between chunks from the same document. maximumPagesToTake limits how many pages are returned, so setting it to 1 drops the rest of the document, and an overlap of 0 means no overlapping text. degreeOfParallelism is a Custom Web API skill setting for parallel calls, not a chunking control.
Deep dive
How it works
The Text Split skill breaks text into chunks, either sentences or pages of a configurable length. In pages mode, the skill tries not to cut a sentence, so a page can be a little shorter than the maximum. Even so, a key idea can still land across a page boundary. That's what pageOverlapLength is for.
In pages mode, pageOverlapLength makes each page start with that number of characters or tokens from the end of the previous page. Context near a boundary then appears in both chunks, so retrieval is less likely to return half a thought. If it's set to 0, successive pages don't overlap.
Related pages-mode settings:
- maximumPageLength: the maximum page size. In characters it ranges from 300 to 50,000, with a default of 5,000. For text embedding models, the general recommendation is about 512 tokens.
- unit (preview): chunk by characters (the default) or by
azureOpenAITokens. This affects both maximumPageLength and pageOverlapLength. - maximumPagesToTake: how many pages to return. The default of 0 returns all pages.
The other options, and when you'd use them
- pageOverlapLength set to 0: An overlap of 0 means no overlapping text, which is the cause of the problem, not the fix.
- maximumPagesToTake set to 1: This returns only the first page. Use it only when you need a subset of pages; here it would drop the rest of each document.
- Increase degreeOfParallelism: This is a Custom Web API skill setting for how many calls the indexer makes in parallel to your endpoint, from 1 to 10 with a default of 5. It affects throughput, not chunk boundaries.
Exam tip
"Sentences split across chunks" or "chunks lack context" means adding pageOverlapLength in pages mode. "Chunks too big for the embedding model" means lowering maximumPageLength.
References
Last reviewed: