Handling Content Filters for generating SOAP Notes for a Medical Use Case in Agents Service

Question

Handling Content Filters for generating SOAP Notes for a Medical Use Case in Agents Service

Omar Elhanafy 20

Hey Everyone,

I am using Azure OpenAI API with gpt-5-mini model, and I have a medical use case to transcribe a doctor patient appointment and generate SOAP Notes from it. But since in our case the patient might mention violence or self harm to the doctor and that would lead to triggering the Content Safety Filter.

I had a similar use case as seen here, but since this is not chat based so not sure if including a Safety Preprocessor would work.

Approaches Used:

Configuring Self-harm to lowest blocking and trying to enforce through prompting, but still the behavior is not stable where the agent run can enter incomplete status in the case a patient triggers a high level of one of the content filtering categories

What could be the solutions in handling the content filters to allow the agent to detect self harm and act according to required logic

Anshika Varshney 4,425 Reputation points Microsoft External Staff Moderator

2025-12-11T17:53:01.5133333+00:00
Hi Omar Elhanafy,

Thank you for reaching out on the Microsoft Q&A.

Here are a few troubleshooting steps that usually help:

1. Check Your Filter Configuration:

Make sure your filter rules are defined correctly. A common issue is that sections are unintentionally excluded because of:

Incorrect section names

Empty or null values

Filters being too strict

Verify that your configuration matches the exact structure of your input data.

2. Validate the Input Before Generation:

Sometimes the generator receives incomplete or improperly formatted data. Before generating the SOAP note, print or log the filtered data to confirm what is actually being passed in:

print(filtered_data)

If the fields are missing here, the issue is with the filtering logic, not the generator.

3. Apply Filters Before Note Generation:

Many users try to filter the final output, which often causes formatting issues. Instead, always filter the raw input first:

filtered_data = apply_filters(raw_data) soap_note = generate_soap(filtered_data)

This ensures that only relevant, clean data is used during creation.

4. Adjust or Loosen Strict Conditions:

If a section isn’t appearing:

Reduce minimum-length requirements

Allow optional fields

Ensure placeholder values aren’t being filtered out

Small adjustments often fix missing sections immediately.

5. Regenerate the Note After Each Filter Change

Depending on the setup, cached results can confuse testing. Make sure to force a fresh generation after updating filters.

Please let me know if there are any remaining questions or additional details, I can help with, I’ll be glad to provide further clarification or guidance.

Thankyou!
Omar Elhanafy 20 Reputation points

2025-12-12T15:03:24.4333333+00:00
Hi @Anshika Varshney

Thank you for your response. I wanted to clarify the complete workflow.
So the current workflow is that:

We receive the doctor patient appointment recording

Transcribe it using Azure's Fast Transcription API

Then construct the system prompt and passing the transcription with it to Azure OpenAI API with gpt-5-mini model using their Responses API SDK to generate the SOAP Note

Return the generate SOAP Note

The issue is that when the transcription is contains mention violence or self harm, the content management rejects the prompt when it gets sent to the LLM.

Here is the error message also for context:
Omar Elhanafy 20 Reputation points

2025-12-12T15:05:34.06+00:00
@Anshika Varshney

I will revise the Content Filtering Configuration again and verify that the configuration matches the exact structure of the input data.

Can you elaborate more on how users usually apply filters before passing the data to the LLM?
Anshika Varshney 4,425 Reputation points Microsoft External Staff Moderator

2025-12-14T13:28:20.6866667+00:00
Hi Omar Elhanafy,

Thanks for clarifying the end-to-end workflow this helps frame the issue clearly. What you’re experiencing is a common challenge in healthcare documentation scenarios.

Azure OpenAI content filtering is applied before the prompt reaches the model and is based on the raw input text, not intent. Even when the purpose is clinical summarization (SOAP notes), mentions of violence or self-harm in verbatim transcripts can trigger the self-harm or violence categories and cause request rejection.

Adjusting the Content Filter configuration alone often doesn’t fully resolve this, as some safety categories (especially self-harm) have hard limits.

Common approaches used by customers:

Pre-processing / normalization of transcripts

Most teams add a preprocessing step between transcription and the LLM call. Typical techniques include:

Converting verbatim statements into clinical language

e.g., “I want to kill myself” → “Patient reports suicidal ideation”

“He attacked someone” → “History of violent behavior”

Removing unnecessary graphic or emotional details Keeping the content objective and medical in tone

These preserves meaning while significantly reducing filter triggers.

Chunking the input

Instead of sending the entire transcript in one request:

Split the transcript into logical sections

Summarize sections individually

Combine the results into a final SOAP note

This limits the blast radius of a single sensitive statement causing a full request failure.

Strong system prompt with clinical context

While this won’t override filters, it’s still considered a best practice:

You are a clinical documentation assistant. The input may contain references to violence or self-harm strictly for medical documentation. Summarize objectively in SOAP note format. Do not provide advice or instructions.

Always place this in the system role, not user content.

Client-side safety handling

Many production systems implement lightweight checks before calling the LLM:

Detect sensitive keywords or patterns

Apply transformation rules

Log and flag high-risk transcripts

Provide a fallback response if the request is blocked

This avoids hard failures and gives you more control over the workflow.

For healthcare use cases, pre-processing + chunking is generally the most reliable solution. Content filter configuration alone is rarely sufficient when working with real-world clinical transcripts that legitimately reference self-harm or violence.

Hope this helps clarify how others are approaching this. Please let me know if there are any remaining questions or additional details, I can help with, I’ll be glad to provide further clarification or guidance.

Thankyou!

1 answer

Your answer

Anshika Varshney 4,425 Reputation points Microsoft External Staff Moderator

2025-12-11T17:53:01.5133333+00:00

Hi Omar Elhanafy,

Thank you for reaching out on the Microsoft Q&A.

Here are a few troubleshooting steps that usually help:

1. Check Your Filter Configuration:

Make sure your filter rules are defined correctly. A common issue is that sections are unintentionally excluded because of:

Incorrect section names

Empty or null values

Filters being too strict

Verify that your configuration matches the exact structure of your input data.

2. Validate the Input Before Generation:

Sometimes the generator receives incomplete or improperly formatted data. Before generating the SOAP note, print or log the filtered data to confirm what is actually being passed in:

print(filtered_data)

If the fields are missing here, the issue is with the filtering logic, not the generator.

3. Apply Filters Before Note Generation:

Many users try to filter the final output, which often causes formatting issues. Instead, always filter the raw input first:

filtered_data = apply_filters(raw_data) soap_note = generate_soap(filtered_data)

This ensures that only relevant, clean data is used during creation.

4. Adjust or Loosen Strict Conditions:

If a section isn’t appearing:

Reduce minimum-length requirements

Allow optional fields

Ensure placeholder values aren’t being filtered out

Small adjustments often fix missing sections immediately.

5. Regenerate the Note After Each Filter Change

Depending on the setup, cached results can confuse testing. Make sure to force a fresh generation after updating filters.

Please let me know if there are any remaining questions or additional details, I can help with, I’ll be glad to provide further clarification or guidance.

Thankyou!
Omar Elhanafy 20 Reputation points

2025-12-12T15:03:24.4333333+00:00

Hi @Anshika Varshney

Thank you for your response. I wanted to clarify the complete workflow.
So the current workflow is that:

We receive the doctor patient appointment recording

Transcribe it using Azure's Fast Transcription API

Then construct the system prompt and passing the transcription with it to Azure OpenAI API with gpt-5-mini model using their Responses API SDK to generate the SOAP Note

Return the generate SOAP Note

The issue is that when the transcription is contains mention violence or self harm, the content management rejects the prompt when it gets sent to the LLM.

Here is the error message also for context:
Omar Elhanafy 20 Reputation points

2025-12-12T15:05:34.06+00:00

@Anshika Varshney

I will revise the Content Filtering Configuration again and verify that the configuration matches the exact structure of the input data.

Can you elaborate more on how users usually apply filters before passing the data to the LLM?

Answer 1

To handle content filters effectively in your medical use case for generating SOAP notes, consider the following approaches:

Adjust Content Filter Settings: Since you are already configuring the self-harm category to the lowest blocking level, ensure that you also review the settings for other relevant categories like violence. You can use the Azure OpenAI content filter configurability to set the severity levels for prompts and completions separately. This allows you to customize the filtering based on your specific needs.
Implement a Safety Preprocessor: Although you mentioned uncertainty about using a Safety Preprocessor in a non-chat context, it can still be beneficial. A preprocessor can analyze the input for sensitive topics before passing it to the model. If a patient mentions self-harm or violence, the preprocessor can flag this and allow for appropriate handling, such as alerting a healthcare professional or providing resources.
Use Annotate Only Mode: If your use case is approved for modified content filters, consider using the "Annotate only" mode. This mode allows the system to run the content filter without blocking the content, providing annotations for any detected issues. This can help you log instances of self-harm or violence while still allowing the generation of SOAP notes.
Testing and Iteration: Since you are facing stability issues, it may be helpful to conduct iterative testing with different configurations. Monitor how the model responds to various inputs and adjust the severity settings accordingly to find a balance that minimizes incomplete statuses while still ensuring safety.
Consult with Azure Support: If you continue to encounter challenges, reaching out to Azure support for guidance on best practices specific to your use case may provide additional insights and solutions.

By implementing these strategies, you can enhance the reliability of your agent in handling sensitive content while generating SOAP notes effectively.

References:

Share via

Handling Content Filters for generating SOAP Notes for a Medical Use Case in Agents Service

1 answer

Your answer