Quick question for prompt setup.

bbmjcoding

Quick question for prompt setup.
How would I go about wording the prompt template to set a specific context?
i.e when the user asks "How do I turn off an instance?" it can't respond based on the indices. But if the query is worded "How do I turn off an instance in AWS" it will generate a response. So I would need the prompt template to always assume the user queries are relating to AWS.

9 comments

DDeleted User

one idea is to prefix the prompt:

Plain Text

const userInput = "how do i turn off an instance?"

const promptPrefix = "assume the following input is related to AWS:"

const prompt = `${promptPrefix} ${userInput}`;

console.log(prompt);

vanilla mistral gives me:

Plain Text

To stop an Amazon Elastic Compute Cloud (Amazon EC2) instance in Amazon Web Services (AWS), follow these steps:

1. Sign in to the AWS Management Console and open the Amazon EC2 console at https://console.aws.amazon.com/ec2/.
2. In the navigation pane, choose "Instances."
3. Select the check box next to the name of the instance you want to stop. The Actions pane appears on the right side of the page.
4. Choose "Instance State > Stop Instances" from the Actions dropdown list.
5. In the Stop Instances dialog box, confirm that the correct instance is selected and review any associated warnings or notes. You can also 
choose an option for how to handle any associated Elastic IP addresses and IMDSv2 instance metadata options.
6. Choose "Yes, Stop Now" to stop the instance. The instance's state changes to "shutting down." Once the instance is no longer running, it will
be in the "stopped" state and you will not be charged for any usage.

DDeleted User

could extend beyond AWS with something like:

Plain Text

const promptPrefix = `assume the following input is related to ${provider}:`

where provider is some state value

bbmjcoding

I'm using the qa_prompt_tmpl so it would be slightly different than this

bbmjcoding

qa_prompt_str = (
"Context information is below.\n"
"---------------------\n"
"{context_str}\n"
"---------------------\n"
"Given the context information and not prior knowledge, "
"answer the question: {query_str}\n"
)

DDeleted User

yes this is the same concept of prefixing the query (sorry JS is my goto!)

just above your block of code, i would modify query_str like this:

Plain Text

query_str = f"assume the following input is related to AWS: {query_str}"

or, if you want to make it more dynamic like above:

Plain Text

context_str = "AWS"

query_str = "how do i turn off an instance?"

prompt_prefix = f"assume the following input is related to {context_str}:"

qa_prompt_str = f"{prompt_prefix} {query_str}"

bbmjcoding

Looks like that's working, but not necessarily giving me the same answer as "How do I turn off an instance in AWS". It gives an answer, but not as detailed/precise. Here's what I did:

Plain Text

context_str = "AWS"
prompt_prefix = f"Assume the following input is related to {context_str}:"
qa_prompt_str = 
"You are User assistant. Context information is given below. \n"
"------------\n"
f"{prompt_prefix}\n"
"------------\n"
"Given the context information and not prior knowledge, "
"answer the query in detailed manner and point wise if required. \n"
"Query: {query_str}\n"
"Answer: "
)

DDeleted User

Have you tried using the exact example I gave? The reason I say this is because in my (limited) experience with using different models, sometimes smalltalk in prompts kinda gets in the way. I wonder if it's getting hung up on some of this prompt's design

DDeleted User

in a perfect world with a perfect model, your prompt is ideal, it's logical and well laid out, but I wonder if the LLM is interpreting it all

DDeleted User

just giving f"{prompt_prefix} {query_str}" as the prompt

Add a reply

Find answers from the community

Quick question for prompt setup.