how to fix inconsistencies in Content Understanding API Response?

Question

I have been using Azure AI Content Understanding to extract data from financial statements.

I am seeing inconsistent behavior between Content Understanding Studio and the Runtime API.

For the same document and the same analyzer:

• Analyzer Test in Content Understanding Studio consistently returns the expected result.
• Runtime API calls sometimes return a different result and extract data from sections that should be excluded.

Questions:

1. Is Analyzer Test in Content Understanding Studio guaranteed to use the exact same inference pipeline as the Runtime API?
2. Is Content Understanding deterministic for the same document and analyzer?
3. Can API executions return different extraction results for the same document and analyzer?
4. Can an operation reach "Succeeded" before all extraction artifacts are fully finalized?
5. Are there any known differences between Studio testing and Runtime API execution?

Any guidance would be greatly appreciated.

Thank you.

Answer 1

Hello Celeste Ifran,

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

I understand that you would like to know how you can fix inconsistencies in Content Understanding API Response.

The issue is withing a Runtime execution mismatch or missing extraction-boundary enforcement, such as different API version, model deployment mapping, analyzer build, input document, missing page range, or accepting extracted fields without validating their source and spans. The documented Runtime API flow returns final extracted fields only after the operation reaches Succeeded, so Succeeded should not be treated as an incomplete state. - https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/quickstart/use-rest-api, https://learn.microsoft.com/en-us/rest/api/contentunderstanding/content-analyzers/get-result?view=rest-contentunderstanding-2025-11-01

To resolve this, you will need to:

• Use the GA Content Understanding API version 2025-11-01.
• Explicitly pass the same model deployments in the Runtime API request instead of relying on resource defaults.
• Use the input range property to exclude pages that must not be analyzed.
• Enable returnDetails and estimateFieldSourceAndConfidence.
• Accept extracted values only when their source, spans, and confidence prove they came from the approved financial statement section.
• If valid and excluded sections are on the same page, use classification or segmentation with contentCategories and route only the allowed section to the extraction analyzer.
• If the analyzer still selects the wrong section, add corrected labeled samples in Studio and rebuild the analyzer before retesting through the Runtime API. - https://learn.microsoft.com/en-us/rest/api/contentunderstanding/content-analyzers/analyze?view=rest-contentunderstanding-2025-11-01, https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/concepts/models-deployments, https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/concepts/analyzer-reference, and https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/document/analyzer-improvement gives you more insights.

After the Runtime API is pinned to the same API version, same model deployments, same analyzer, same document bytes, and the extracted fields are validated against source locations, the API will no longer silently accept values from excluded financial statement sections. If inconsistent results still occur after this controlled setup, collect the Operation-Location, operationId, x-ms-request-id, x-ms-client-request-id, full request JSON, and result JSON, then escalate to Azure Support/Product Group for backend trace comparison. Use the associated resource links for more reading and implementation steps.

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

---

Please do not close the thread by upvoting and accepting the answer if any part of it is helpful.