GPT 5.4 in the deployment always responded with "I'm sorry, but I cannot assist with that request."

Question

Everything works fine previously, but today the model responses with "I'm sorry, but I cannot assist with that request." even when user says "hi". There are no code changes or system prompt changes.

If this is an content filtering issue, can we disable it for all? Currently we still have to apply for 4 required by Microsoft.

Answer 1

Hi Xinyu He

You can find co-relation id from SDK call.

client = OpenAI(
    base_url=endpoint,
    api_key=api_key
)

# 🔹 Use with_raw_response to access headers
response = client.chat.completions.with_raw_response.create(
    model=deployment_name,
    messages=[
        {
            "role": "user",
            "content": "What is the capital of France?"
        }
    ],
)

# ✅ Chat completion content
completion = response.parse()
print("Answer:", completion.choices[0].message.content)

# ✅ HTTP headers (case-insensitive)
headers = response.headers

print("\n--- Azure OpenAI Request Metadata ---")
print("APIM Request ID:", headers.get("apim-request-id"))
print("Correlation ID:", headers.get("x-ms-correlation-request-id"))
print("Request ID:", headers.get("x-request-id"))
print("Region:", headers.get("x-ms-region"))
print("Date:", headers.get("date"))

Thank you for confirming that you were able to mitigate the issue by disabling web search tool.

From the ICMs, Issue seems coming from Content filtering control plane, Severity threshold was set to high for content filters for GPT-5 series models.

Thank you for your inputs.