.Net: Bug: Hallucination in semantic kernel responses

[Vishal11848]

Describe the bug
We have integrated the semantic kernel and configured it to call the ticketing system (API based) automatically by using an auto-function call. During the initial function call, it retrieves the response accurately and provides the correct answer. However, when follow-up questions are asked within the same chat history, it starts generating random answers and exhibiting hallucinations. Additionally, it is not even hitting the function again on the follow-up question. If the chat history is cleared and a new conversation is started, it performs correctly initially, but the same issues reoccur after 2nd or 3rd question.

Expected behavior
It should give an answer correctly, and call function every time instead of hallucinat.

Platform

• OS: Windows
• IDE: Visual Studio
• Language: C#,
• Source: Microsoft.SemanticKernel (1.15.0), Microsoft.SemanticKernel.Connectors.OpenAI (1.15.0)

[markwallace-microsoft]

Hi @Vishal11848, take a look at this article Managing Chat History for Large Language Models (LLMs). You likely need to implement one of these strategies to limit the amount of chat history being sent to the LLM.

[Vishal11848]

Hi @Vishal11848, take a look at this article Managing Chat History for Large Language Models (LLMs). You likely need to implement one of these strategies to limit the amount of chat history being sent to the LLM.

Thank you for your suggestion, @markwallace-microsoft . Currently, we are sending the history of the past five conversations. However, the system sometimes starts hallucinating as early as the 2nd interaction, and other times it occurs at the 3rd or 5th interaction. The issue is not consistent.

I am trying to follow the steps which you shared, will share more updates on it.

[Vishal11848]

Hello @markwallace-microsoft

Could you please inform me about the version of the semantic kernel being used here?

[sphenry]

@Vishal11848 what model are you using? Does the issue reproduce with more advanced models like GPT-4o

[Vishal11848]

Hello @sphenry yes, this issue is reproducible with GPT-4o

[markwallace-microsoft]

@Vishal11848 this isn't an issue specific to Semantic Kernel, it's a well known LLM issue, so the version of Semantic Kernel won't be a factor. Our general guidance is always use the latest version of Semantic Kernel, as it will contain the most up-to-date fixes.

The general guidance for reducing hallucinations is:

1. Provide clear and specific prompts. Provide relevant context in the prompt (or via function calling) to ground the LLM. It also helps to include links in the grounding data and ask the LLM to provide citations.
2. Use active mitigation i.e. change temperature or frequency_penalty values to modify the behaviour of the LLM.
3. Use multi-short prompting i.e. provide examples of the type of results you expect the LLM to return.

Another thing to try is review your system prompt e.g. it may help to provide boundaries for the LLM to use when generating responses.

For example, if I set my system prompt to You are an AI assistant that helps people answer question about the Python programming language. For all other questions please politely decline to answer. while using gpt-4o-mini and ask Who is the greatest soccer player of all time? the LLM will respond with I'm here to help with questions about Python programming. If you have any questions related to Python, feel free to ask!.

So a suitable system prompt may be what you need to keep the LLM responses relevant and prevent halluncinations.