If this is answered elsewhere, please let me know and provide links.
We would like to have full control over the model used in a Copilot Studio project, with particular request for GPT-4 32k and other higher performance / context models. We have been able to build custom copilots in Azure AI Studio with really great results, but every time we test in Copilot Studio the responses are weak and we cannot figure out how to get the same performance as we are seeing in AI Studio
If we can use any Azure OpenAI models, we'd also like to know if we can use non OpenAI models such as Llama and others with Copilot Studio.
Thanks
Regarding the difference in quality, it's important to note that the current connection between Copilot Studio and Azure Open AI is limited to Semantic Search only. If you have been using a hybrid model (vector + semantic search) in your lab, you might observe some reduced effectiveness in the results.
If you want to have complete control over your model (be it GTP3.5, GPT4, Mystral, or any other), you have the option to replace the generative answer node with an API call to your model. This can be done through Power Automate or directly via the HTTP action in Copilot Studio.
In this scenario, if you're using a traditional generative AI model (input/output), you will need to handle the conversation history. On the other hand, if you're using Azure Open AI or Open AI assistants, there's no need to worry about managing the conversation history since it's handled at the Thread level.
I hope it helps !
Kr,
Y.
Still no answer on this? Copilot Studio does not give me true Copilot capabilities unless I can control the underlying model and model parameters (temp, etc.). Every time my team or my clients try to apply Copilot Studio as a generative AI solution the results are useless.
That is correct. GPT-4 32k, GPT-4 Turbo and others with control over temp and size of prompts and completions, etc.
Basically being able to select from the Azure OpenAI models that are available - https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models - would be the primary ask.
I believe that jltreads was asking if he could use the OpenAI GPT-4 LLM to replace the rather simple NLU model which is the default. He needs richer completions and a far bigger context window.
Good morning,
I hope you are doing well.
Are you using your model on your company data ? If this is the case, maybe the problem doesn't stem from you model but from your indexing. When testing models in Azure AI studio, by default you use keyword search for your indexing but in copilot studio, the default is "simple" which works kinda weirldy (https://learn.microsoft.com/en-us/azure/search/query-simple-syntax) (and the complete documentation https://learn.microsoft.com/en-us/rest/api/searchservice/preview-api/search-documents ).
In order to improve results, the best way to improve results is to use semantic search in azure ai search (https://learn.microsoft.com/en-us/azure/search/semantic-how-to-enable-disable?tabs=enable-portal ).
If you want to use non openAI models, a workaround would be to use the http node of copilot studio in order to call the API of the model.
I hope it helps.
Have an amazing day.
WarrenBelz
146,751
Most Valuable Professional
RandyHayes
76,287
Super User 2024 Season 1
Pstork1
66,091
Most Valuable Professional