Seeking advice from community experts on Copilot Studio Gen AI vs. Copilot for M365
I have created a Copilot Studio bot with Gen AI capability enabled. I manually uploaded multiple purchasing agreements / contracts into the Copilot Studio bot (Content Moderation set to Medium), and also the same set of documents into my Copilot for M365 interface.
When I submit the same prompt into both interface (e.g. What are the standard payment term in all the documents?), I will get different results, with the ones from my Copilot Studio chatbot much more limited.
Even for the same prompt and data source, for Copilot Studio chatbot, the responses seem to always come from only one or two documents that Gen AI selects, while Copilot for M365's are usually more extensive.
Are there any technical workarounds / functionality / prompt writing techniques to make my Copilot Studio chatbot more "intelligent" and comprehensive, especially in being able to process and compare/summarize multiple documents for the same prompt?
Any advice would be helpful, at least for my understanding on the limitations of Gen AI in Copilot Sutdio.
I've experienced similar issues with Copilot Studio struggling to provide more extensive answers based on the context and content of my documents/knowledge, particularly with complex structures.
One approach that improved its performance for me was creating better prompts. Additionally, I found that using the "Topic" function to divide complex documents and knowledge into different sections or topics helped significantly. This approach seems to aid the model in chunking the documents and understanding the content more effectively. It is also possible to set up triggers and prompts for these topics, which can help the model locate the relevant content more accurately.
This might be a current limitation of Copilot compared to models like GPT-4 or Gemini, which tend to handle complex documents more effectively. I assume this may have to do with the models used in Copilot and the context window limitations of the current model, making it challenging to perform Retrieval Augmented Generation (RAG) processes to provide better answers.