Azure AI Search - Text Split Skill related

Hi @G Sanil

Thanks for your clear and detailed question about the Text Split skill in Azure AI Search — it's a great one, and you're spot on with your observations.

Quick answers to your questions:

1. Yes — when textSplitMode is set to sentences, the skill splits the text so that each sentence becomes its own individual chunk. It breaks strictly on sentence-ending punctuation (., ?, !, etc., respecting the language set in defaultLanguageCode).
   
2. Correct — parameters like maximumPageLength, pageOverlapLength, and maximumPagesToTake are ignored in sentences mode. They only apply when textSplitMode is set to pages. The skill simply performs a clean punctuation-based split in sentences mode.
   
3. Recommended use cases for sentences mode:
   • You need very fine-grained, precise chunks where preserving exact sentence boundaries is important.
   • Your content consists of short, self-contained sentences and you want natural language units rather than fixed-length chunks.
   • You're doing highly targeted semantic search or analysis where one sentence = one meaningful retrieval unit.

That said, for most vector search / RAG scenarios, Microsoft recommends using pages mode instead. It gives you better control over chunk size (to stay within embedding model token limits) and supports overlap for improved context. Sentences mode often creates a much larger number of chunks, which can increase indexing time and storage costs.

Reference :

https://learn.microsoft.com/en-us/azure/search/cognitive-search-skill-textsplit

https://learn.microsoft.com/en-us/azure/search/vector-search-how-to-chunk-documents

Please do not forget to click "Accept the answer” and Yes, this can be beneficial to other community members.

If you have any other questions, let me know in the "comments" and I would be happy to help you

Hi, G Sanil

When using the Text Split Skill in Azure AI Search, the behavior of the textSplitMode parameter changes significantly depending on whether you choose pages or sentences. The differences you observed are expected, and the parameters you mentioned are only applicable in specific modes.

Below is a clear breakdown of how the skill behaves and how each mode is intended to be used.

1. Does textSplitMode = sentences create one chunk per sentence?

Yes. When textSplitMode is set to sentences, the skill splits the input strictly along sentence boundaries. Each sentence becomes its own chunk, and the skill does not attempt to merge or size-balance them.

This mode is designed to preserve natural linguistic boundaries rather than enforce token or character limits.

2. Are maximumPageLength, pageOverlapLength, and similar parameters ignored in sentence mode?

Correct. All page‑related parameters apply only when textSplitMode = pages.

When using sentences mode:

• maximumPageLength is ignored

pageOverlapLength is ignored

maximumPagesToTake is ignored

The skill does not attempt to enforce chunk size or overlap when splitting by sentences.

3. Recommended use cases for textSplitMode = sentences

Sentence‑level splitting is useful when you need:

Highly granular semantic units

Fine‑grained embeddings for downstream retrieval

Precise alignment between input text and model responses

Scenarios where each sentence may carry independent meaning (for example, FAQs, short statements, or structured content)

This mode is typically chosen when the goal is to maximize semantic precision rather than optimize for chunk size or token efficiency.

Summary

pages mode is for controlled chunking with size limits and overlaps.

sentences mode is for natural linguistic segmentation with no size‑based controls.

The parameters you listed are intentionally ignored in sentence mode.

If you share more about your indexing or retrieval scenario, I can help recommend which mode would be more effective.

Hope this helps, Martin

I look forward to your update and am happy to continue working with you until the issue is resolved. If you find the answer helpful, please click "Accept Answer" and consider upvoting it. Otherwise, please keep me posted by clicking "Add comments" below instead of selecting Yes or No.

This response is drafted with the help of AI systems.