Unanswered

GPT Model for creating generative AI answer

(1) Share

Report

Posted on by Yang-Liu

Hi, can I ask what GPT (3.5?) or LLM Models does the Create generative answers use in the Copilot Studio? I've been reading through the Microsoft documentation but couldn't find any details. Thank you.

Categories:

General topics

Create generatvie...

I have the same question (0)

All responses (8)

Answers (0)

HenryJammes on at

Like (2)

Report

Hi @Yang-Liu,

May I ask why you're looking for this information?
Today, generative answers are using gpt-35-turbo for the text summarization, but the model gets updated when newer ones outperform them on a number of dimensions and become available. For more complex tasks, like creating a topic with copilot, it's using gpt-4.

Henry

Was this reply helpful? Yes No
Yang-Liu 17 on at

Like (1)

Report

Hi, thank you very much for your quick response. We're building some demos using copilot studio and want to have a better understanding on the generative answer feature.

Was this reply helpful? Yes No
bhatiavinay99 5 on at

Like (1)

Report

Thanks for sharing currently used model numbers (as of Jan/Feb 2024). Is this documented anywhere in Copilot Studio documentation.

For real world deployments, the LLM require to be validated and approved for use by internal Information Security or equivalent teams before being used in production to answer employee or customer queries with Generative AI boost capabilities.

What level of control will tenant admins have to manage what version of LLM model is used, when the model should upgrade from GPT3.5 to GPT4.0? Will it follow Microsoft Business Apps Release cadence?

One option is to connect to Azure Open AI which gives us more control.

Was this reply helpful? Yes No
HenryJammes on at

Like (2)

Report

As a SaaS offering, as part of generative answers, this is abstracted to our customers, but as you're mentioning, you can choose not to use this feature and establish your own connection with a LLM and be responsible of the desired updates and guardrails. For generative answers, not only do we update a model (e.g. gpt-35-turbo) but we also take care of updating the model version (e.g. 0125). We also ship improvements to the meta prompts and moderation layers. I'd love to understand more about what the desired experience should be regarding these?

Was this reply helpful? Yes No
JuhaniR 2 on at

Like (1)

Report

@hehe Thanks for the information on the used model. We plan to use Copilot Studio in Finnish language and in my own tests GPT4-turbo outperforms GPT3.5-turbo by a large margin. It is able to reason more accurately and the Finnish language responses are more grammatically correct. It also produces much less hallucinationa in it’s responses.

Do you have plans on updating the model you use to GPT4 in the near future or offering an easy way for changing the model?

One option would be to use Azure OpenAI studio connection and use GPT4 through that. We will mostly use Sharepoint as a data source and out of the box Copilot Studio seems to be pretty good at searching through our Sharepoint to find relevant information. Is it possible to pass the search results snippets that Copilot Studio has found based on the users requst to the GPT4-model connected through Azure OpenAi Studio? Or do we need to connect our Sharepoint data to the model in Azure OpenAI studio separately?

Was this reply helpful? Yes No
bhatiavinay99 5 on at

Like (1)

Report

Thanks for the response. We discussed the desired experience with your colleagues from Microsoft Product Group over a call last week. We suggested to start by publishing the currently used model numbers with built in LLM. And offer admins some control to first upgrade and test in Dev/Test environments before upgrade the LLM version in Production.

Was this reply helpful? Yes No
ArneBernadotte 2 on at

Like (1)

Report
I am uncertain about the final conclusion in this thread.

Having used GPT-4 for a year and being well-acquainted with its capabilities, my initial attempts to create a copilot for an internal document today suggest the use of an inferior and practically useless LLM for our purposes.

Are there any plans to:
Clearly document which LLMs are being used?
Allow users to directly select the LLM (such as GPT-4) within the Copilot Studio application, without needing to connect to Azure Open AI?

Was this reply helpful? Yes No
Mahesh Chintha 158 on at

Like (1)

Report

@HenryJammes

Copilot studio still using gpt-35-turbo for generative answers and gpt-4 for topic creation? Are there any changes/upgrades to the model versions being used? Thank you!

Was this reply helpful? Yes No