Azure OpenAI studio chat playgorund with own data

Faisal Saleem 20 Reputation points
2024-01-06T15:43:23.08+00:00

Hi,

I am very new to AI and am doing some discovery work for a customer on what benefit they might get using AI services.

The question i have is around the chat playground when using your own data, I have a sql db that has service desk ticketing system type data i.e ticket number, subject, description, category etc

I have added this data to the Azure search service and created indexes etc and then pointed the playground at it using one of the gpt models.

What i want to know is a bit more insight as to how this works because it seems it does generate answer based on the entire dataset, it seems more like it randomly chooses a number of documents (created by the indexer process i guess) and uses this to provide answers. this also means that any questions asking for counts do not work correctly either.

for example if i have a 1000 rows in the db this translates to approx 1000 documents and if i ask the question like how many tickets are open, i would get the answer like 4 or 10 where as there are hundreds of open tickets.

So the question is, is this expected behavior or have a missed something or some configuration that is preventing this?

do i have the wrong expectation and it simply cannot do this?

The same responses occur when i ask something like 'what is the most common ticket category' it will give me an answer but based on only a subset of the entire dataset, is there some sort of limitation in place causing this?

thanks

Azure OpenAI Service
Azure OpenAI Service
An Azure service that provides access to OpenAI’s GPT-3 models with enterprise capabilities.
3,623 questions
0 comments No comments
{count} votes

Accepted answer
  1. Boris Von Dahle 3,131 Reputation points
    2024-01-07T13:26:35.2+00:00

    Hello,

    When you add your SQL database to Azure Search, the service indexes your data. This index acts as the source from which the AI model retrieves information to answer queries. The quality and structure of this index significantly influence the model's performance.

    When you use a GPT model in the Azure AI chatbot, the model generates responses based on the indexed data it can access.

    AI models, especially those like GPT, may not be inherently designed for specific analytical tasks such as counting the number of open tickets. Their strength lies in natural language understanding and response generation, rather than performing database-like count operations.

    In order to improve the accuracy ensure that your data is indexed in a way that makes it easily retrievable and understandable by the AI model. This might involve structuring the data or metadata in a certain way.

    Also try to modify how your queries are made to ensure it retrieves and uses the most relevant data from your indexes.

    For specific tasks like counting or categorizing, consider using AI solutions that are tailored for analytical operations, possibly in combination with the GPT model.

    Hope this helps
    Regards

    If you found this answer useful, please consider marking it as 'Accepted.' This helps other users easily find and benefit from this information


0 additional answers

Sort by: Most helpful

Your answer

Answers can be marked as Accepted Answers by the question author, which helps users to know the answer solved the author's problem.