We are using the prebuilt-tax.us model to process W-2 forms. The field "LocalWagesTipsEtc" is present in the document, but the model is not extracting it in the result.

Hello Lokesh Saini,

Welcome to Microsoft Q&A and Thank you for reaching out.

If you’re using the prebuilt-tax.us model to process W-2 forms and fields like LocalWagesTipsEtc (under LocalTaxInfos) are present in the document but not appearing in the result, this is typically due to one of the following reasons:

1.Field Behavior & Confidence Threshold

Local tax fields (W-2 Boxes 18–20) are:

• Optional

Often repeated (multiple localities)

Returned only when the model has sufficient confidence

If the OCR confidence is low (small font, light print, skewed scan, overlapping text), the model may detect the text but not populate the structured field.

Please verify,

• Check the raw OCR output (content, words) to confirm the value was read.
• Review confidence scores for nearby fields.

2.Array Handling

LocalTaxInfos is returned as an array, not a single object.

If your code expects a single object instead of iterating through the array, it may appear as if the field is missing.

Make sure you are checking:

documents[0].fields["LocalTaxInfos"]

and iterating through all entries.

3.Tier Limitation

If you are using the F0 (free) tier, only the first two pages are analyzed.

If your W-2 content extends beyond those pages, the local tax section may not be processed. Upgrading to S0 removes this limitation.

4.Supported Field Validation

Confirm that LocalWagesTipsEtc is included in the supported schema for the W-2 model version you are using. While it is generally supported, schema support can vary by API version.

Also ensure you are using the latest available Document Intelligence API version, as updates may improve extraction coverage.

5.Layout Variations

The prebuilt model is trained on common IRS W-2 layouts. Extraction may fail if:

The employer uses a non-standard template

The local tax section is shifted or compressed

There are alignment inconsistencies

In such cases, text may appear in OCR but not be mapped to the structured field.

6.Alternative steps

If the prebuilt model does not reliably extract the required fields:

Try the General Document model to extract key-value pairs.

Consider training a custom model if you frequently process a consistent W-2 layout that differs from standard formats.

Custom models can improve accuracy for repeated document structures.

Recommended Steps

Verify the document quality (clear scan, high resolution).

Confirm you are not hitting F0 tier page limits.

Check that you are iterating through LocalTaxInfos as an array.

Review the raw OCR output to confirm text detection.

Confirm API version and SDK are up to date.

If the issue persists with clean samples, test in Document Intelligence Studio for comparison.

If reproducible across multiple valid W-2s, open a support request with:

A redacted sample document

  API version
  
     Full JSON response
     

This will help determine whether it is a parsing issue, confidence threshold behavior, layout variation, or a model extraction limitation.

Please refer this

• Document Intelligence US tax document models
• Diagnosing Unextracted Fields

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

Thank you!

minutes will recheck...

so.... hi Lokesh Saini,

this is usually not a bug but a model coverage limitation. The prebuilt-tax.us model does not guarantee extraction of every possible W-2 field, especially local sections, because those fields vary heavily by state and employer formatting. The model is optimised for the most standard federal boxes, and local tax sections are often inconsistent in layout and labelling.

See the model version u are using. are u on the latest API version and model release, because field coverage sometimes expands between versions. Review the raw layout output using the prebuilt-layout model on the same document. If the text for LocalWagesTipsEtc is detected correctly in layout but missing from prebuilt-tax.us structured fields, then it confirms it is a schema coverage limitation rather than an OCR issue.

Whether the field is actually present in the JSON schema returned by the model. If LocalTaxInfos exists but is empty, the model likely did not reach sufficient confidence to populate it. In that case try testing multiple W-2 samples to see if extraction works for some formats but not others. If consistent extraction of local tax fields is critical, the recommended approach is to build a custom model using Azure Document Intelligence custom extraction trained specifically on ur W-2 variants. Prebuilt models are generalised, but custom models are far more reliable for structured but variable sections like local tax boxes.

If needed, open a support request with sample redacted W-2s so msft can confirm whether that specific field is currently supported in the prebuilt-tax.us schema.

rgds,

Alex