Azure AI Document Intelligence
An Azure service that turns documents into usable data. Previously known as Azure Form Recognizer.
Custom Classification - Inconsistent results between Doc Studio "Test Model" and API
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(16, 45%, 11%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EJH%3C/text%3E%3C/svg%3E)
Jack Halpern
20
Reputation points
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.
Please advise. Thanks.
Jack
Azure AI Document Intelligence
{count} votes
-
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(316.8, 39%, 40%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ESM%3C/text%3E%3C/svg%3E)
Sridhar M 2,525 Reputation points Microsoft External Staff Moderator2025-12-02T23:31:45.85+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, andclassificationConfidenceThreshold. If your API request omits or changes these, classification can differ.
- The Studio often uses recommended defaults for
- 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, andautoRotateif applicable. - Confirm Model Version Check the
modelIdand 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!
- API Parameters vs. Studio Defaults
-
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
localeandimageEnhancement(internally mapped toOcrHighResolution) are API-only settings. They are not exposed in the Studio UI for custom extraction testing.Why they’re missing in
ClassifyDocumentOptions-
ClassifyDocumentOptionsis 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
AnalyzeDocumentOptionsin 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);-
DocumentAnalysisFeatureenum maps to the Studio toggles:- High resolution →
OcrHighResolution - Style font →
StyleFont - Formulas →
Formulas - Barcodes →
Barcodes - Languages →
Languages
- High resolution →
Sign in to comment
Sign in to answer