update

Author: mrbullwinkle

articles/ai-services/openai/faq.yml

@@ -109,9 +109,15 @@ sections:
           How do I fix InternalServerError - 500 - Failed to create completion as the model generated invalid Unicode output?
         answer:
           You can minimize the occurrence of these errors by reducing the temperature of your prompts to less than 1 and ensuring you're using a client with retry logic. Reattempting the request often results in a successful response.
+      - question: |
+          How do I fix Server error (500): Unexpected special token
+        answer: |
+          This is a a known issue. You can minimize the occurrence of these errors by reducing the temperature of your prompts to less than 1 and ensuring you're using a client with retry logic. Reattempting the request often results in a successful response. 
+          If reducing temperature to less than 1 does not reduce the frequency of this error an alternative workaround is set presence/frequency penalties and logit biases to their default values. In some cases, it may help to set `top_p` to a non-default, lower value to encourage the model to avoid sampling tokens with lower probability tokens.  
+          Another way to solve this issue is to add logit_bias in the payload. For example, if error is `Unexpected special token: 100266`, then we can add "logit_bias": {100266:-100} to reduce the issue by using token 100266 mentioned in the service side error message. 
       - question: |
           We noticed charges associated with API calls that failed to complete with status code 400. Why are failed API calls generating a charge? 
-        answer:
+        answer: 
           If the service performs processing, you will be charged even if the status code is not successful (not 200).
           Common examples of this are, a 400 error due to a content filter or input limit, or a 408 error due to a time-out. Charges will also occur when a `status 200` is received with a `finish_reason` of `content_filter`.
           In this case the prompt did not have any issues, but the completion generated by the model was detected to violate the content filtering rules, which result in the completion being filtered.