Power Platform Community Forum Thread Details

Does GPT fetch SharePoint Files by their file names and not content?

In my chatbot use case, my user asks about a form, and the bot is supposed to find the file in the SharePoint "shared documents". The process fails. let's have this example:

There is a file titled "request for equipment check.docx"
The file is a simple form, with "name, ID, office address, etc." data fields. In the tile, nor in the body of the form, the name of the form is not mentioned. In other words, you cannot see any "request for equipment check" word inside the file.
the user asks: "I need to check my equipment. is there a form I can fill?"
the bot fails to return the link.

Is it even possible for the GPT in Copilot Studio to fetch a file (as a reference) by its title? In order to rule out other possibilities:

Bot's access to the document location is verified. Sometimes it can actually reference "some" file but not the right one.
for other forms that have the keyword inside the form, "sometimes" it works. for example. if the form is "work-related complaint", and the title is repeated in the header, or there is a data field "main complaint"

I can assume there must be trick solutions like maintain a list of title / link, and define it as a knowledge source; but I'm more interested to know the capability of the copilot studio, and hopefully solve this puzzle without hacks.

Categories:

General

Hello,

in general the files and content on the Sharepoint site is indexed, which means that probably (I'm not 100% how it works as I don't have the access to the codebase :-)) everything is converted to embeddings - it's a set of vectors for the machine to understand what the data is like. Then when you prompt the model:

1. It converts your prompt to an embedding.

2. It searches the knowledge that it has to find the nearest neighbour, based on your prompt embedding.

3. It outputs it, if it finds anything.

Given those facts it all depends on how you prompt it, so it won't be that stable as users can use various prompts. The best way for now to make it more stable is to use a different datasource. In your case you could do something like:

1. Define a SP List with rows containing the title, description and a link to the file.

2. Use generative orchestration with action Get items and instruct the model to fill ODATA filter, based on the user prompt.

2. Use classic approach and create a topic with trigger phrases, questions and Get items action to fetch the data.

That way the bot will be able to filter out the data first, before presenting it to the user.

In case of any other questions, let me know. If the answer helped you, mark it, so that others can benefit from it.

Best regards,

Artur Stepniak

Quick Links

Leaderboard