Azure OpenAI Service
An Azure service that provides access to OpenAI’s GPT-3 models with enterprise capabilities.
2,645 questions
I encountered issues with the accuracy and reliability of responses generated by the Azure OpenAI service.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(233.6, 86%, 32%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EAS%3C/text%3E%3C/svg%3E)
Ashlesha Sharma
0
Reputation points
I have been using the Azure OpenAI service for a while now, and I have encountered several issues with the responses generated by the system. These issues have led to inaccuracies and inconsistencies in the information provided, impacting the overall user experience.
The task is to generate a few tags on the basis of some given information. Using prompts the code is hitting Azure's OpenAI API to generate tags in response but got weird results:
- Getting a set of weird instructions, which might be executed to generate results.
- Response is including different languages in response.
3 Getting all results as 'UNKNOWN'
- Instead of having enough token left, the response is not complete and truncated by '....'
In order to investigate this further, I decided to conduct a comparison by using the same inputs in both the Azure OpenAI service and the OpenAI-Chat GPT service. The results were quite surprising. The responses generated by the OpenAI-Chat GPT service were consistently more accurate and aligned with the intended meaning of the queries.
I'm seeking help here. Thanks in advance.
{count} votes
-
romungi-MSFT 43,696 Reputation points Microsoft Employee2023-07-13T08:19:15.4766667+00:00 @Ashlesha Sharma Is the inconsistency still seen with responses? Could you also add the prompts used when these responses were generated.
-
Ashlesha Sharma 0 Reputation points
2023-07-13T10:16:39.88+00:00 Thanks romungi-MSFT for responding, The following is the prompt used:
You are a food expert who can provide accurate details on food item based on the following recipe details. Set A : GIVEN_ATTRIBUTES: 1.recipe_id: unique id for recipe 2.recipe_name: The name of the recipe 3.ingredients: ingredient used to make it with respective quantity. 4.instructions : The step by step cooking process of this recipe. Set B: The following are the set of attributes with description and the possible values, which you need to figure out for each shared recipe details: 1.preparation_time: The time taken to prepare the recipe. 2.meal_period: The Time when the recipe should be served like: Lunch / Dinner / Breakfast / Others 3.cooking_method: The methods of cooking 4.ingredient_origin_type: The source or place of origin of a particular ingredient, indicating where it comes from or is produced example: "Local / Imported / Organic / GMO-free / Others" 5.cooking_tool: Cooking Tools invloved to prepare the recipe. 6.staple_ingredients: The one or more essential ingredient that forms the basis of a diet or meal. Provide the values for Set B item to following recipes: recipes = {recipes_list} Instructions for creating response: 1. Provide accurate value for Set B attribute by analyzing Set A attributes. Avoid answering 'Others' or 'NA', instead try to provide a value for it. 2. For cooking tools and cooking methods, refer cooking instructions to prepare answer. 3. One Attribute-cell can hold multiple values seprated by '/' only , if there exists more then one value for respective attribute for the given recipe. 4. The output should be a valid json array where each property name should be enclosed in double quotes and also include the recipe_id for each json object. 5. Provide complete result, do not leave it as unterminated string. -
romungi-MSFT 43,696 Reputation points Microsoft Employee
2023-07-18T05:43:36.89+00:00 @Ashlesha Sharma I have tried using your prompt and it seems to generate consistent responses as mentioned in the instructions. I have used the playground to lookup the response though.
Sign in to comment