Unanswered

SharePoint Knowledge Source gives hallucinations? Looking for advice

(3) Share

Report

Posted on by MM-07030830-0

Hey all, I'm using Copilot Studio and have connected a SharePoint site as a knowledge source. The site hosts around 1800 PDF documents, and I've added the entire site as a single knowledge source for my agent.

The problem is that the agent frequently gives factually incorrect answers, even when the correct information clearly exists in the documents; it often "hallucinates" and/or mixes up details. I'm assuming it's trying to parse too much unstructured data at once.

What I did:

turned on orchestration (alongside generative AI) and turned off "Allow the AI to use its own general knowledge"
turned on enhanced search results (yes, I do have Microsoft 365 Copilot license in the same tenant as the agent)
added the content moderation to be medium (maybe should try with high as well now)

Possible next steps:

I saw in some documentation that SharePoint has a limit of a total of 200 files which can be included for each source, so maybe I should break my knowledge sources into 9 separate sources containing 200 PDF documents each. (docs: Unstructured data as a knowledge source - Microsoft Copilot Studio | Microsoft Learn). But, I wonder if this is relevant in my case since my understanding is that there are two types of SharePoint knowledge sources (SP URL, and direct documents from SP).

Wondered if anyone has any tips on what is the best way to move forward? I understand that most likely I will need to enhance the description of the knowledge source and maybe clean the source some more (some folders are unnecessary and can be removed). Would it help if I added some sort of semantic mapping in the general instructions? What I mean is that I explain where to look for (relative path) in the knowledge source in case a certain question is being asked.

Any tips or opinions I would greatly appreciate. :)

Categories:

General topics

I have the same question (0)

All responses (6)

Answers (1)

ronaldwalcott 3,847 Super User 2025 Season 2 on at

Like (0)

Report

What type of queries are generating hallucinations? Is it possible that the data contains conflicting information? Could also be caused language phrasing.

Was this reply helpful? Yes No
Jcook 7,781 Most Valuable Professional on at

Like (1)

Report

Hi,

you can try a few things:

1. I would try with Enhanced Search turned off.

2. Try adding retrieval hints in the agents instructions.
example:

- For payroll-holiday questions, prefer files in /HR/Benefits/Holidays.

Was this reply helpful? Yes No
MM-07030830-0 39 on at

Like (0)

Report

@Jcook, any explanation as to why I should turn off Enhanced Search? The second suggestion I am already implementing at the moment of writing - will let you know after testing in a couple of days.

@ronaldwalcott, the problem comes from general questions on the knowledge. The answers are present in the documents and there are no conflicting information (tough to be 100% sure because of the big count of files, but I analyzed a lot of the documentation and I haven't found any conflicting information).

Was this reply helpful? Yes No
sabe 3 on at

Like (0)

Report

I have the same situation, a large SP doc library with company policies, procedures, and I cannot get precise answer consistently regarding how to create an out of office message. In the knowledge document, there is exact text to be included in the ooo message, but the agent is unable to quote it word by word. It misses and replaces words, phrases.

Even worse when I turn on Enhanced Search, so I suggest to turn it off.

Was this reply helpful? Yes No
SB-26110046-0 3 on at

Like (0)

Report

My agent is generating responses using its own general knowledge with GenAI Orchestration Enabled and General Knowledge Disabled.

I have a single knowledge source which is a word document uploaded directly to the agent. I ask a question that has no answer in the document, and it generates an answer anyway.

GenAI Orchestration seems broken when it comes to knowledge.

Was this reply helpful? Yes No
Verified answer

MM-07030830-0 39 on at

Like (0)

Report

What gave me better results in the end was a lot of data cleaning (removing duplicate files in knowledge source) and breaking each folder into a separate SharePoint knowledge source.

Additionally, ingesting files directly as unstructured data gave me way better responses.

Hopefully some might find this useful :)

Was this reply helpful? Yes No