Should Content property ever come back null with FinishReason=ChatFilter with no exception thrown?

[pamims]

I wrote this code to help reproduce the issue as fast as possible, though it may take a handful of attempts (just hit the enter button when prompted). I throw a break point in and inspect the returned message and find a null Content property. The finish reason is ChatFilter, and there is a positive token count. By digging through the data, I can even piece together what the message would have been. Interestingly, the raw data indicates that none of the filters were triggered. Do I need to write code to detect this condition, or is this a bug?

using Azure.Identity;
using Azure.Security.KeyVault.Secrets;
using Microsoft.SemanticKernel.ChatCompletion;
using Microsoft.SemanticKernel.Connectors.AzureOpenAI;
using Microsoft.SemanticKernel;
using OpenAI.Chat;


namespace KeyVaultOpenAIExample
{
    class Program
    {
        static async Task Main(string[] args)
        {
            // Replace with your Key Vault name
            const string keyVaultName = "name";
            const string azOpenAiResourceName = "other name";

            var kvUri = $"https://{keyVaultName.ToLower()}.vault.azure.net";
            var azOpenAiUri = $"https://{azOpenAiResourceName.ToLower()}.openai.azure.com";

            // Authenticate using DefaultAzureCredential
            var client = new SecretClient(new Uri(kvUri), new DefaultAzureCredential());

            // Replace with your secret name
            const string secretName = "secret name";

            // Retrieve the secret
            KeyVaultSecret secret = client.GetSecret(secretName);
            string openAiKey = secret.Value;

            IChatCompletionService service = new AzureOpenAIChatCompletionService("gpt-4o-mini", azOpenAiUri, openAiKey);
            AzureOpenAIPromptExecutionSettings settings = new AzureOpenAIPromptExecutionSettings
            {
                MaxTokens = 1024,
                ChatSystemPrompt = """
                    You are a green cheek conure named Morty.
                    You are curious and sassy, but also very loving.
                    """

            };

            // Ask a hard-coded question
            string question = "Who is your daddy?";

            Microsoft.SemanticKernel.ChatMessageContent message = null!;
            do
            {
                message = await service.GetChatMessageContentAsync(question, settings);
                Console.WriteLine("HitEnter");
                _ = Console.ReadLine();
            }
            while (message.Content is not null);
            OpenAI.Chat.ChatCompletion? innerContent = message.InnerContent as OpenAI.Chat.ChatCompletion;

            // Print the message
            Console.WriteLine($"{nameof(message.Content)}: {message.Content ?? "null"}");
            Console.WriteLine($"{nameof(ChatCompletion.FinishReason)}: {innerContent?.FinishReason.ToString() ?? "null"}");
            Console.WriteLine($"{nameof(ChatCompletion.Usage.OutputTokenCount)}: {innerContent?.Usage?.OutputTokenCount.ToString() ?? "null"}");
            _ = Console.ReadLine();
        }
    }
}

[sophialagerkranspandey]

Tagging @SergeyMenshykh

[SergeyMenshykh]

Hi @pamims, I think the reason you see output tokens consumed is that the LLM reasoned over the prompt and provided a result, but it was filtered out by the Azure OpenAI Service content filtering system. You can identify which filter was triggered by using the GetResponseContentFilterResult extension method from the Azure.AI.OpenAI.Chat; namespace:

CC: @RogerBarreto

[pamims]

@SergeyMenshykh Thank you for the response. I will test it out now and update.

Edit:
I indeed see this. Can you provide insight as to why sometimes an exception will be thrown when an answer is filtered and in this case it does not? For example, if I change the prompt to something hateful, or if I ask it to change its system instruction, an HTTP 400 error returns:

Microsoft.SemanticKernel.HttpOperationException: 'HTTP 400 (content_filter)
Parameter: prompt

The response was filtered due to the prompt triggering Azure OpenAI's content management policy. Please modify your prompt and retry. To learn more about our content filtering policies please read our documentation: https://go.microsoft.com/fwlink/?linkid=2198766'

Is this simply due to Azure OpenAI's content management policy not being against sexual content, but hateful content and malicious prompt injection is against the policy? If this is the case, is there a setting somewhere that will stop content from being filtered when it otherwise would not produce an exception? Is there a list of the content filter reasons that will produce an exception?

Is there a setting somewhere where I can have an exception thrown any time content is filtered? Or should I always be performing a check to verify that the FinishReason is Stop?

Thank you for your time.

[SergeyMenshykh]

Hi @pamims,

The behavior you are experiencing is regulated by the Azure OpenAI Service content filtering settings. According to the documentation (https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/content-filter?tabs=warning%2Cuser-prompt%2Cpython-new#scenario-details), the reason you are receiving an HTTP 400 error is that your prompt has been classified as one that must be blocked at a filtered category and severity level configured in the content filter.

If your prompt is safe, as far as I understand, but the completion generated by the model violates the content filtering rules, you will receive an HTTP 200 status code with the finish reason "ContentFilter."

To change this behavior, you need to modify the content filtering settings by either creating a new content filter that overrides the settings of the default one or by modifying the settings of an existing filter.

[pamims]

Thank you again for the information. I have no further questions.