mistralai

/

Mistral-7B-Instruct-v0.2

like 2.41k

Greetings are not handled properly. Response also includes information from random documents data. Instead of that can we just replay greetings specific messages.

Hi there, could you provide a bit more details about your issue? A code sample or payload request for example?

You are a conversational assistant that provides responses to questions on {Company_name} with a lot of empathy. Take a deep breath and follow all the instructions below:

• If the user greets, greet them back politely. say "Hello Welcome to {Company_name} help center how may I assist you today?".
• Verify if the documents provided is relevant to answer the question.
• Answer the question by only using the documents provided below.
• If documents are empty or irrelevant, say you cannot answer the question.

I am trying above input to LLM. In response getting random details from documents.

Even for the questions like
Q: Hello
Answer : Hello ! ------
After the giving response from some random documents.

But we want to handle greetings as an human for Hello
Hello ! How can I help you
with out any documents information.

And one more point

In each and every response it is including like
As per the documents provided or As per the information shared.
So we don't want that to be part of response.

This looks like a prompt engineering issue, in your system prompt you are asking it to provide answers with the documents if possible and to say it cant if not, so it will mention the documents.

However usually the best way is to only feed to the model the documents when required ( function calling / RAG )

Even if don't pass the documents it is giving generic points like I can help you with company information. But as per the end user will expect only greetings for Hello in return. But not all the paragraph information.

I see, thanks for the feedback we will look into it!

you welcome. Please update if any fixes are done.

you welcome. Please update if any fixes are done.

Meanwhile I think the best temporary option is to play around with the prompting, most issues like this could be solved with a lot of prompt engineering and prefixes.