Answered

The agent alters my input. For example, 'Bae Hyun-deok' becomes 'Bae Hyeong-deok'

(0) Share

Report

Posted on by CU11060340-1

"Hello. I am currently building an internal RAG chatbot using Copilot Studio. My knowledge sources include Excel files, Docx documents, and websites. I have categorized them into six distinct topics and configured the 'Generative Answers' node to ensure the bot references only the relevant knowledge base for each query.

However, I am encountering an issue: when using GPT-4o, the reasoning capability feels slightly limited, though the answers are accurate. When I switch to GPT-5 (or higher-tier models), the bot starts modifying the output text—changing specific terms or names, such as 'Bae Hyun-deok' becoming 'Bae Hyeong-deok'.

Could you explain why this text distortion occurs? I am relatively new to this field, so a detailed explanation would be greatly appreciated."

Categories:

Building Copilot Studio chatbots in Microsoft Teams

Model context protocol

Topic creation & management

I have the same question (0)

All responses (3)

Answers (1)

Sort by

Suggested answer

11manish 2,837 on at

Like (0)

Report
Copy link

Link copied!

The text distortion is occurring because the model is generating an answer from the retrieved content rather than simply copying it. More advanced models such

as GPT-5 tend to normalize, correct, or reinterpret text—especially names, acronyms, and technical terms.

To reduce this behavior in Copilot Studio, strengthen your grounding instructions, explicitly tell the model not to modify names or identifiers, enable citations,

and use structured data sources for critical business information that must be reproduced exactly.

Was this reply helpful? Yes No
Verified answer

chiaraalina 2,318 Super User 2026 Season 1 on at

Like (0)

Report
Copy link

Link copied!

Hi @CU09061234-0

I think it is because GPT-5 is doing more interpretation. That helps for reasoning tasks, but it can hurt when you need exact text. That is useful when the question is complex, but it also means the model may try harder to “understand” and “clean up” the answer. Instead of copying a name exactly, it may produce the version that looks most likely or most natural.

I had a similar problem and fixed it with a Custom Search and a Prompt after in a Topic.

Let me know if you have more questions!

Was this reply helpful? Yes No
CU11060340-1 4 on at

Like (0)

Report
Copy link

Link copied!

@chiaraalina

Hello. If you don't mind me asking, could you explain where and what exact phrasing you entered for the prompt, and how you organized the topics when there were multiple files? I would really appreciate it if you could let me know."

Was this reply helpful? Yes No