arugyani
commented
3 months ago Preliminary Prompts
Contributors: @prajwalpkn @arugyani
Using guidance from Prompt Engineering Guide, the following prompts were developed and tested using Claude, GPT-3.5, and GPT-4. The prompts and outputs shown below are our current iterations after some hours working and iterating on various prompt techniques.
Key Notes
- Prompts for generating Q/A info and selecting filters should be separated to reduce prompt complexity, allowing the LLM to be more consistent.
- We are defining output format by the LLM at the bottom of our prompts. Need to look into as to whether there is a way to guarantee this output format.
- Should include an informational blurb in the prompt.
- Multiple Answers should be allowed for certain questions to increase diversity when doing product search
Prompts & Sample Outputs
Claude 3 Prompt
Basic Prompt
As an expert in cameras, generate a question to guide a customer in choosing the right camera product. Provide 4 multiple-choice options, each no longer than 5 words, to help the user make their purchasing decision. Use the user's response to the question to determine the next most relevant question to ask. Indicate whether the user can select multiple answers for the question.
Format the output in JSON as follows:
{ "question": "Your question here", "multipleAnswers": true, "options": [ { "label": "A", "text": "Option 1, max 5 words" }, { "label": "B", "text": "Option 2, max 5 words" }, { "label": "C", "text": "Option 3, max 5 words" }, { "label": "D", "text": "Option 4, max 5 words" } ] }
Continue this JSON format for up to 6 questions in total, each question based on the user's previous response. Set "multipleAnswers" to true or false for each question to indicate whether multiple answer selections are allowed. Respond with only the JSON.
[
{
"question": "What's your main purpose for the camera?",
"answers": ["Sports photography", "Vlogging", "Travel photography", "General use"],
"multipleAnswers": false
},
{
"question": "Which sports will you be photographing?",
"answers": ["Water sports", "Indoor sports", "Mountain sports", "Racing sports"],
"multipleAnswers": true
},
{
"question": "How important is video quality to you?",
"answers": ["Very important", "Somewhat important", "Not very important", "Unimportant"],
"multipleAnswers": false
},
{
"question": "Do you need a camera that's easy to carry?",
"answers": ["Yes, very light", "Somewhat portable", "Not a priority", "Size doesn't matter"],
"multipleAnswers": false
},
{
"question": "How important is weather resistance?",
"answers": ["Very important", "Somewhat important", "Not very important", "Unimportant"],
"multipleAnswers": false
},
{
"question": "What's your budget range?",
"answers": ["Under $200", "$200-$500", "$500-$1000", "Above $1000"],
"multipleAnswers": false
}
]GPT Prompts & Outputs
Given the context of identifying a user's camera needs, particularly for beginners, create a simplified questionnaire. The goal is to make the process user-friendly while capturing essential preferences and requirements related to camera usage.
### Instructions ###
- Avoid preliminary descriptions. Directly list the questions with their respective multiple choice options.
- Formulate 5-6 questions designed to be easily understandable by someone with minimal knowledge of cameras. Each should aim to uncover specific needs or preferences.
- Provide 4 multiple choice answers for each question. These options should encompass a wide range of common preferences and needs. Keep answers concise, ideally under 5 words.
- Allow for multiple answers on certain questions where users might have more than one preference. Indicate clearly which questions permit multiple selections.
- Following the questionnaire, synthesize the responses to recommend camera specifications that match the identified preferences.
### Questionnaire Format ###
Responses should be structured in JSON format, adhering to the following template:
[
{
"question": "Your question here",
"answers": ["Option 1", "Option 2", "Option 3", "Option 4"],
"multipleAnswers": true or false
}
...
]
Please ensure each question is designed to be approachable and informative, helping users articulate their camera needs without needing in-depth technical knowledge.[
{
"question": "What's your main purpose for the camera?",
"answers": ["Sports photography", "Vlogging", "Travel photography", "General use"],
"multipleAnswers": false
},
{
"question": "Which sports will you be photographing?",
"answers": ["Water sports", "Indoor sports", "Mountain sports", "Racing sports"],
"multipleAnswers": true
},
{
"question": "How important is video quality to you?",
"answers": ["Very important", "Somewhat important", "Not very important", "Unimportant"],
"multipleAnswers": false
},
{
"question": "Do you need a camera that's easy to carry?",
"answers": ["Yes, very light", "Somewhat portable", "Not a priority", "Size doesn't matter"],
"multipleAnswers": false
},
{
"question": "How important is weather resistance?",
"answers": ["Very important", "Somewhat important", "Not very important", "Unimportant"],
"multipleAnswers": false
},
{
"question": "What's your budget range?",
"answers": ["Under $200", "$200-$500", "$500-$1000", "Above $1000"],
"multipleAnswers": false
}
]
prajwalpkn
LLM Prompt Engineering - Questions & Answers
In this issue we test different prompts for getting user details and filter specifications from the LLM.
In addition, we will specify response formats for a consistent result.