Intermittent vectorization failure (HTTP 400) in SharePoint index via AI Foundry during specific time window

[Daniel-Kosyra]

Hi,

I am experiencing an intermittent issue with an Azure AI Search index that is connected to a SharePoint knowledge source and used by an agent deployed via Microsoft AI Foundry (New).

Environment / Setup

• Platform: New Microsoft AI Foundry

• Search Service: Azure AI Search

• Data Source: SharePoint

• Embedder: text-embedding-3-large

• Components configured:

  • Data source (SharePoint)
  • Knowledge store / knowledge base
  • Index
  • Indexer (scheduled daily at 23:00)

• Skillset includes:

  • ContentUnderstandingSkill
  • AzureOpenAIEmbeddingSkill (for text)
  • ChatCompletionSkill
  • ConditionalSkill (for image verbalization)
  • AzureOpenAIEmbeddingSkill (for generated image descriptions)
  • AIServicesByKey

An agent is deployed and configured to use this index as its knowledge source.

Issue Description

Initially, the system worked as expected. However, I have observed a consistent time-based issue:

• Between approximately 08:00 and 13:00 (CET):

  • The agent fails to retrieve knowledge from the index.
  • The index appears to return errors related to vectorization.

• Outside of this time window:

  • The index functions normally.
  • The agent successfully retrieves and uses indexed knowledge.

Error Message

During the affected time period, the following error occurs:

"Could not complete vectorization action. The vectorization endpoint returned status code '400' (BadRequest)."

Observations

• The issue appears time-dependent and recurring daily.
• The indexer schedule (23:00) does not align with the failure window.
• The system works reliably outside the affected hours without configuration changes.
• The failure impacts both indexing behavior and downstream agent performance.

Questions / Clarifications

1. Are there any regional or time-based throttling limits or resource constraints (e.g., Azure OpenAI or vectorization endpoints) that could explain this behavior?
2. Could this be related to capacity limits, scaling behavior, or quota exhaustion during peak hours?
3. Are there known issues with AzureOpenAIEmbeddingSkill or vectorization endpoints returning intermittent 400 errors under load?
4. What diagnostics (logs, metrics) would you recommend to further isolate the root cause?

Request

I would appreciate guidance on:

• Identifying the root cause of this time-based failure
• Recommended monitoring or logging to confirm the issue
• Any configuration changes or scaling recommendations to ensure consistent availability

Please let me know if additional logs or diagnostic information would be helpful.

[0xsan-z]

We are also getting the same behavior and the model is text-embedding-3-small , deployed in uksouth and Deployment type is GlobalStandard. In our case there is no burst of traffic to the embedding model, it is just me doing minimum queries, still get this randomly. If i adjust the model TPM, even lowering TPM value, it starts to work.