Custom Classification - Inconsistent results between Doc Studio "Test Model" and API

Question

Custom Classification - Inconsistent results between Doc Studio "Test Model" and API

Jack Halpern 20

We've been successfully using the Doc Intelligence API to Classify and then Extract fields from documents for the last few months. However, occasionally we have encountered instances where our Custom Classifier fails to correctly classify a document or pages within a multi-page document. While that is not unusual or unexpected when we investigate further, and run the SAME documents through the SAME Custom Classifier using the Doc Studio we get DIFFERENT results - the studio is able to successfully classify while the API fails. It's not as if it's random - if we run multiple tests the Studio always gets the correct result, and the API always gets the same invalid result.

Is this a known issue?

Here's an example. The first screenshot is using the studio and the second screenshot from our API log with both the same document and same classifier. Note the difference between the results. The Studio correctly identifies this as multi-page "checks" while the API result is different.

User's image

Please advise. Thanks.

Jack

Sridhar M 2,525 Reputation points Microsoft External Staff Moderator

2025-12-02T23:31:45.85+00:00
Hi Jack Halpern

Welcome to Microsoft Q&A and Thank you for your question.

This behavior is usually due to differences in configuration and processing context between the Studio and API runtime. While both use the same underlying model, the Studio applies certain defaults and optimizations that may not be mirrored in your API calls.

Common Causes:

API Parameters vs. Studio Defaults

The Studio often uses recommended defaults for pages, locale, and classificationConfidenceThreshold. If your API request omits or changes these, classification can differ.

Document Preprocessing

Studio may apply additional preprocessing (e.g., image enhancement, orientation correction) before classification. API calls require you to enable similar options explicitly.

Version or Deployment Differences

Ensure the API is calling the same model version as the Studio. If the Studio is on a newer version, results can diverge.

Multi-page Handling

Studio automatically aggregates classification across all pages. API calls need explicit handling for multi-page documents; otherwise, page-level classification might skew results.

Recommended Actions:

Match Studio Settings in API Review the Studio’s “Analyze options” and replicate them in your API request (e.g., pages: "1-3", locale, classificationConfidenceThreshold).

Enable Preprocessing Features Use parameters like ocrMode, imageEnhancement, and autoRotate if applicable.

Confirm Model Version Check the modelId and version in both Studio and API calls to ensure consistency.

Log and Compare Inputs Validate that the exact same document bytes and page ranges are sent to the API.

Note: Not a bug in the classifier itself but a configuration mismatch between Studio defaults and API usage. Microsoft recommends aligning API parameters with Studio settings for consistent results.

I Hope this helps. Do let me know if you have any further queries.

Thank you!
Sridhar M 2,525 Reputation points Microsoft External Staff Moderator

2025-12-03T16:27:40.9166667+00:00

Hi Jack Halpern

Did you get any chance to review the above response. Thank you!
Jack Halpern 20 Reputation points

2025-12-03T16:47:55.6566667+00:00

Thank you, yes. I am reviewing your suggestions now. I only see a few options. See screenshot. Where are the other you mentioned? locale, imageEnhancement, etc? I also don't see them in the C# API "ClassifyDocumentOptions" object.
Jack Halpern 20 Reputation points

2025-12-04T15:47:08.4833333+00:00

I see other options but only on the Extraction, not Classification. Is there a comparable option for Classification?

Jack
Sridhar M 2,525 Reputation points Microsoft External Staff Moderator

2025-12-04T19:58:29.3933333+00:00
Hi Jack Halpern

The Analyze options dialog in Document Intelligence Studio only shows a subset of features for the Test workflow of a custom extraction model:

Optional detection: Barcodes, Language

Premium detection: High resolution, Style font, Formulas

Features like locale and imageEnhancement (internally mapped to OcrHighResolution) are API-only settings. They are not exposed in the Studio UI for custom extraction testing.

Why they’re missing in ClassifyDocumentOptions

ClassifyDocumentOptions is for document classification only. It supports:

ClassifierId

Pages

Input source It does not include analysis add-ons or locale hints because classification doesn’t require OCR tuning.

Use AnalyzeDocumentOptions in the C# SDK for analysis scenarios:

var options = new AnalyzeDocumentOptions(modelId, new Uri(pdfUri)) { Locale = "en-US", // language hint Pages = "1-3", // page range Features = { DocumentAnalysisFeature.OcrHighResolution, // imageEnhancement DocumentAnalysisFeature.StyleFont, DocumentAnalysisFeature.Formulas, DocumentAnalysisFeature.Barcodes, DocumentAnalysisFeature.Languages } }; var operation = await client.AnalyzeDocumentAsync(WaitUntil.Completed, options);

DocumentAnalysisFeature enum maps to the Studio toggles:

High resolution → OcrHighResolution

Style font → StyleFont

Formulas → Formulas

Barcodes → Barcodes

Languages → Languages

Your answer

Sridhar M 2,525 Reputation points Microsoft External Staff Moderator

2025-12-03T16:27:40.9166667+00:00

Hi Jack Halpern

Did you get any chance to review the above response. Thank you!
Jack Halpern 20 Reputation points

2025-12-03T16:47:55.6566667+00:00

Thank you, yes. I am reviewing your suggestions now. I only see a few options. See screenshot. Where are the other you mentioned? locale, imageEnhancement, etc? I also don't see them in the C# API "ClassifyDocumentOptions" object.
Jack Halpern 20 Reputation points

2025-12-04T15:47:08.4833333+00:00

I see other options but only on the Extraction, not Classification. Is there a comparable option for Classification?

Jack
Sridhar M 2,525 Reputation points Microsoft External Staff Moderator

2025-12-04T19:58:29.3933333+00:00

Hi Jack Halpern

The Analyze options dialog in Document Intelligence Studio only shows a subset of features for the Test workflow of a custom extraction model:

Optional detection: Barcodes, Language

Premium detection: High resolution, Style font, Formulas

Features like locale and imageEnhancement (internally mapped to OcrHighResolution) are API-only settings. They are not exposed in the Studio UI for custom extraction testing.

Why they’re missing in ClassifyDocumentOptions

ClassifyDocumentOptions is for document classification only. It supports:

ClassifierId

Pages

Input source It does not include analysis add-ons or locale hints because classification doesn’t require OCR tuning.

Use AnalyzeDocumentOptions in the C# SDK for analysis scenarios:

var options = new AnalyzeDocumentOptions(modelId, new Uri(pdfUri)) { Locale = "en-US", // language hint Pages = "1-3", // page range Features = { DocumentAnalysisFeature.OcrHighResolution, // imageEnhancement DocumentAnalysisFeature.StyleFont, DocumentAnalysisFeature.Formulas, DocumentAnalysisFeature.Barcodes, DocumentAnalysisFeature.Languages } }; var operation = await client.AnalyzeDocumentAsync(WaitUntil.Completed, options);

DocumentAnalysisFeature enum maps to the Studio toggles:

High resolution → OcrHighResolution

Style font → StyleFont

Formulas → Formulas

Barcodes → Barcodes

Languages → Languages

Share via

Custom Classification - Inconsistent results between Doc Studio "Test Model" and API

Your answer