Azure AI Search Service Index Creation

Question

Azure AI Search Service Index Creation

Meenakshi Sankar 60

Hello,

I have created a vector search index using the portal (not SDK). I have a few questions on that

If I run the indexer again with the same set of source documents, is it guaranteed that I will get the same emdeddings and the chunking will remain the same?
I noticed that when I followed the same steps to recreate the index, the document count in the index page was different. Why is that so? Is there anyway we can avoid this by using SDK to create index?
What is the default chunking strategy used? Is the documents split at a page level?
What does the option of Semantic Ranking option do?
Is there a way we can confirm the number of chunks created per document?
What are the advantages the SDK based index creation have over portal based index creation?
What is the default chunking strategy for portal based index creation?

Meenakshi Sankar 60 Reputation points

2026-04-22T13:15:58.2866667+00:00

Also, is it possible to change any of the settings (eg chunking strategy ) used for a portal based index creation?
Praneeth Maddali 8,445 Reputation points Microsoft External Staff Moderator

2026-04-22T14:03:32.56+00:00
Hi @Meenakshi Sankar

When you create a vector index using the Azure portal (Import and vectorize data), Azure AI Search uses integrated vectorization. This means the service automatically handles document chunking and embedding generation through a managed indexer and skillset. Because chunking is token‑based and managed internally, rerunning the indexer or recreating the index can result in different chunk counts, even when the source documents don’t change. This is expected behavior for portal‑based ingestion.

The portal wizard doesn’t split documents strictly by page, and chunk size/overlap settings aren’t configurable there. If you need deterministic chunking, consistent document counts, and full control, Microsoft recommends creating the index, skillset, and indexer using REST APIs or SDKs, where these settings can be explicitly defined.

The Semantic Ranking option doesn’t affect indexing or chunking. It’s a query-time feature that re-ranks search results using Microsoft’s language models to improve relevance and can return semantic captions or answers.

Reference:

Integrated vectorization overview: https://learn.microsoft.com/en-us/azure/search/vector-search-integrated-vectorization

Vector search via Azure portal: https://learn.microsoft.com/en-us/azure/search/search-get-started-portal-import-vectors?tabs=storage-access%2Cblob-storage%2Caoai%2Cvectorize-images

Semantic ranking overview: https://learn.microsoft.com/en-us/azure/search/semantic-search-overview

Create an index using SDKs/REST (recommended for production): https://learn.microsoft.com/en-us/azure/search/search-how-to-create-search-index?tabs=portal

Kindly let us know if the above helps or you need further assistance on this issue.

Please "upvote" if the information helped you. This will help us and others in the community as well.
Praneeth Maddali 8,445 Reputation points Microsoft External Staff Moderator

2026-04-23T11:43:37.1433333+00:00

Hi @Meenakshi Sankar

Just following up to see if your issue has been resolved or if you had a chance to review my earlier comment?
Meenakshi Sankar 60 Reputation points

2026-04-23T22:11:45.43+00:00
Hi Praneeth,

A few additional questions. I am a little confused when you mentioned chunking is token‑based for portal based index creation. I took a look at the skillsets and it mentioned the unit as "character" and textSplitMode: pages. I would assume that it would be character based instead of token based. Not sure if I am missing something here

Additionally, can you please advise how using an SDK or API will make a difference? What settings do we need to change to make them predictable?

Another important question I have is on the Semantic Ranking. Which model is Azure using for this? Also, can you also advise during semantic ranking

does the data stays within Azure environment?

what region does the data gets processed?

Regards,

Meena
Praneeth Maddali 8,445 Reputation points Microsoft External Staff Moderator

2026-04-27T08:45:00.35+00:00
Hi @Meenakshi Sankar

Thanks for the follow-up and for pointing out the skillset details — you're correct.

Chunking Clarification

In the portal-created skillset, the Text Split skill uses textSplitMode: "pages" and unit: "characters" (default). Chunking is therefore character-based, with a default maximumPageLength of 5000 characters. The skill tries to break at sentence boundaries, so actual chunk sizes can vary slightly.

My earlier reference to "token-based" was about the broader integrated vectorization pipeline, which is designed to keep chunks within typical embedding model token limits. However, the actual splitting logic you see is character-based. Minor variations in chunk count can still occur with the portal wizard due to its managed, non-configurable nature.

Making Results Predictable with SDK/REST

The portal wizard doesn't let you control chunk size or overlap, which can lead to slight differences on re-runs.

Using SDKs (.NET, Python, etc.) or REST APIs solves this by letting you define fixed settings in the skillset.

This makes chunking consistent and repeatable with the same source documents. You can also use unit: "azureOpenAITokens" (preview) for token-aware splitting or enable index projections to make each chunk a separate document (easier to count).

Advantages: Full control, version control, reproducibility, and advanced options like Document Layout skill.

Semantic Ranking

It uses Microsoft's proprietary language understanding models (not a public OpenAI/GPT model) to rerank results for better relevance.

Data stays within Azure? Yes. Microsoft does not use your documents, queries, or any customer data from Azure AI Search to train or improve models.

Region for processing? Semantic ranker is available only in selected regions. Processing respects your search service's geography — data stays within the chosen Azure geography (with high-availability replication inside the same geography in most cases).

Semantic ranking is a query-time feature only and does not affect indexing or chunking.

Reference :

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

https://learn.microsoft.com/en-us/azure/search/search-how-to-integrated-vectorization?tabs=prepare-data-storage%2Cprepare-model-aoai

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

https://learn.microsoft.com/en-us/azure/search/semantic-search-overview

Kindly let us know if the above helps or you need further assistance on this issue.

Please "upvote" if the information helped you. This will help us and others in the community as well.
Sina Salam 28,526 Reputation points Volunteer Moderator

2026-04-27T14:55:45.0166667+00:00
Hello Meenakshi Sankar,

Welcome to the Microsoft Q&A and thank you for posting your questions here.

I understand that you have issues with Azure AI Search Service Index Creation.

Do not rely on the wizard run itself for repeatability; inspect or explicitly define the generated index, skillset, indexer, and index projections, set chunking and projection behavior deliberately, and treat index document count as “search documents/chunks,” not “source files.”- You can use the following links for your references: https://learn.microsoft.com/en-us/azure/search/search-import-data-portal, https://learn.microsoft.com/en-us/azure/search/search-how-to-integrated-vectorization, https://learn.microsoft.com/en-us/azure/search/search-how-to-define-index-projections, and https://learn.microsoft.com/en-us/azure/search/search-how-to-manage-index

However, to answer your question one after the other:

If I run the indexer again with the same set of source documents, is it guaranteed that I will get the same emdeddings and the chunking will remain the same?

NO. Not guaranteed (chunking variability), because embeddings models (e.g., text-embedding models) are deterministic per input though chunk boundaries may change, preprocessing may vary, indexer pipeline can: re-tokenize differently and re-split content based on limits.

I noticed that when I followed the same steps to recreate the index, the document count in the index page was different.

Due to chunking and processing differences

Why is that so? Is there anyway we can avoid this by using SDK to create index?

SDK improves control but not magic fix.

What is the default chunking strategy used?

Text Split Skill (token/length-based)

Is the documents split at a page level?

Not strictly page-based

What does the option of Semantic Ranking option do?

Semantic ranking = search relevance only

Is there a way we can confirm the number of chunks created per document?

Not directly visible; must query/debug

What are the advantages the SDK based index creation have over portal based index creation?

SDK gives you full control and reproducibility

What is the default chunking strategy for portal based index creation?

Same as your question number 4.

I hope this is helpful! Do not hesitate to let me know if you have any other questions or clarifications.

Please don't forget to close up the thread here by upvoting and accept it as an answer if it is helpful.
Praneeth Maddali 8,445 Reputation points Microsoft External Staff Moderator

2026-04-28T12:43:40.8233333+00:00

Hi @Meenakshi Sankar

Just following up to see if your issue has been resolved or if you had a chance to review my earlier comment?