Azure Content Understanding Citations and Confidence Score for Generate Fields.

Question

Azure Content Understanding Citations and Confidence Score for Generate Fields.

Siddharth Mehta 1 Microsoft Employee

In Azure Content Understanding, we can extract fields based on Schemas and specify the generation methods by Explicitly setting the method (extract, generate, or classify) for each field based on its purpose.

For Extract fields we can get the Confidence Score, Citation links based on bounded box regions mapping to the document Source

Is the same Now, also Available for Generate fields i.e. getting Confidence Score and Citations for Generate Fields?

Also do we have any code samples for following:

Get confidence score and citations for generate fields using C# SDK
How to Open a PDF and show bounded box regions and highlight Extract (or Generate) fields. Which NuGet package to use for suchPDF operations?

Thanks In Advance!

Anshika Varshney 10,315 Reputation points Microsoft External Staff Moderator

2026-04-20T04:51:32.65+00:00
Hi Siddharth Mehta,

Thanks for your question, this is a good one and something many people are checking now.

Yes, confidence score and citations are available for generate fields as well, but there are a couple of things to keep in mind.

Earlier, confidence and citations were mainly used for extract fields. Now they are supported across different methods including generate, as long as you enable it in your configuration.

So if you want to get confidence score and source locations for generate fields, you need to turn on this option in your analyzer setup.

You can do it in two ways:

Enable it for all fields Set estimateFieldSourceAndConfidence to true

Or enable it at field level Set estimateSourceAndConfidence to true for specific fields

Once this is enabled, the response will include:

The generated value Confidence score Source or citation details like page number and bounding region

So technically, generate fields can show where the value is coming from and how confident the model is.

Now about visualization or highlighting in PDF

The service itself does not highlight anything inside the PDF. It only gives metadata like coordinates and page number.

If you want to show highlights on the document, you need to use a PDF library in your application and draw the bounding boxes using those coordinates.

There is no specific required package mentioned, so you can use any PDF library that supports drawing shapes on documents.

For troubleshooting, if you are not seeing confidence or citations:

Check if the flags are enabled in your analyzer Make sure you are reading the full response JSON Verify that the fields are properly defined in your schema.

Below is a simple example of how the JSON response may look:

{ "fields": { "TotalAmount": { "value": "500 USD", "confidence": 0.92, "source": { "pageNumber": 1, "boundingRegions": [ { "polygon": [100, 200, 400, 200, 400, 250, 100, 250] } ] } }, "GeneratedSummary": { "value": "Payment received for invoice", "confidence": 0.87, "source": { "pageNumber": 1, "boundingRegions": [ { "polygon": [120, 300, 500, 300, 500, 360, 120, 360] } ] } } } }

If these are not enabled, the response will only show values without confidence or source information.

In short:

Generate fields do support confidence and citations But you must enable it explicitly in configuration And for highlighting, you need to handle it in your application using the returned coordinates.

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

Thankyou!
SRILAKSHMI C 18,205 Reputation points Microsoft External Staff Moderator

2026-04-22T04:49:12.71+00:00

Hi @Siddharth Mehta

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank you!

1 answer

Your answer

SRILAKSHMI C 18,205 Reputation points Microsoft External Staff Moderator

2026-04-22T04:49:12.71+00:00

Hi @Siddharth Mehta

We haven’t heard from you on the last response and was just checking back to see if you have a resolution yet. In case if you have any resolution please do share that same with the community as it can be helpful to others. Otherwise, will respond with more details and we will try to help.

Thank you!

Answer 1

Confidence scores and citations (grounding) for generate fields:

Confidence scores are available for all field extraction methods for documents, including generate. This is explicitly called out as a recent enhancement: “Confidence scores are now consistently available across all extraction methods for documents—whether you use extract, generative, or classify.”
Grounding (citations to source regions) is supported for extract and generate methods when enabled. Grounding is defined as the “specific regions in content where a value was extracted or generated,” and is available when confidence/grounding is opted in.
To enable confidence and grounding for field extraction in a document analyzer, set:
- estimateFieldSourceAndConfidence = true in the analyzer config (for all fields), or
- estimateSourceAndConfidence = true on a specific field.

This applies equally to generate fields in document analyzers, so generate fields can return both confidence scores and grounded source regions.

C# SDK sample for confidence and citations on generate fields:

The context describes the API behavior and configuration flags but does not include a C# SDK code sample. The key configuration needed in any client (including C#) is:

In the analyzer schema/configuration for documents:
- Set estimateFieldSourceAndConfidence = true to opt in to confidence and grounding for all fields, or
- On individual fields, set estimateSourceAndConfidence = true.

Once enabled, the JSON response for analyzed documents will include, for each field (including generate fields):

The field value
Confidence score (0–1)
Grounding/source spans (page numbers, bounding boxes, etc.)

The C# SDK would need to deserialize these properties from the service’s JSON response according to the SDK’s model types, but the exact C# types and usage are not provided in the context.

Opening a PDF and showing bounding boxes / highlights:

Content Understanding provides grounding source information for document fields, including page numbers and bounding boxes for extracted values. This is intended to “aid the user experience by highlighting locations for human review and correction.”
The service itself does not perform PDF rendering or visual highlighting. It only returns metadata (page numbers, bounding boxes, spans) that can be used by a client application.
The context does not specify any particular .NET or NuGet package for opening PDFs and drawing bounding boxes/highlights. Any PDF rendering library that allows drawing rectangles at given coordinates could be used, but no specific package recommendation is provided in the available documentation.

References:

Share via

Azure Content Understanding Citations and Confidence Score for Generate Fields.

1 answer

Your answer