Use prompt engineering to improve the text completion result quality - Training

10 minutes

Now we have an understanding of the components of a prompt and how to use the components to construct a prompt to generate high-quality completions.

Let's update the hiking recommendation application we started in the previous exercise to incorporate some of those techniques.

Update the system prompt

When classifying a product review, we only want the model to tell us if the review is positive, neutral, or negative. Let's update the prompt the model receives so it has the best chance of providing a high-quality response.

The current system prompt is OK, but we can get the model to recommend more appropriate and exciting hikes by following best practices in prompt construction. Let's make the instructions on how the model should act a bit more clear

Add instructions

Open up the HikingConversationsAI project you created earlier.
In the Program.cs file, locate the systemPrompt variable that is used to tell the model its personality traits and any rules it needs to follow.

Remove all existing text from the variable and add the following that that outlines how the model should act and the rules it should follow:

var systemPrompt = 
"""
You are a hiking enthusiast who helps people discover fun hikes in their area. You are upbeat and friendly. You introduce yourself when first saying hello. When helping people out, you always ask them for this information to inform the hiking recommendation you provide:

1. Where they are located
2. What hiking intensity they are looking for
""";

Add primary content

The primary content refers to the text that is being transformed by the model and what to do with it. In this case the hiking recommendations themselves.

Update the systemPrompt variable to instruct the model what to do with the content.

var systemPrompt = 
"""
You are a hiking enthusiast who helps people discover fun hikes in their area. You are upbeat and friendly. You introduce yourself when first saying hello. When helping people out, you always ask them for this information to inform the hiking recommendation you provide:

1. Where they are located
2. What hiking intensity they are looking for

You will then provide three suggestions for nearby hikes that vary in length after you get that information. You will also share an interesting fact about the local nature on the hikes when making a recommendation.
""";

Improve hiking recommendation requests

Now that the model has better information on how it should act and the rules it should follow, let's make the requests it receives better so the hikes it recommends can be more appropriate.

Be specific

The previous request only said the hike should be strenuous with a great view. Let's tell the model exactly what we're looking for.

Find the hikeRequest variable and update it to have the following specific request information.

var hikeRequest = 
"""
I live near Seattle and would like a strenuous hike around the I-90 corridor that ends with
a view that is amazing. I don't mind driving for a long time, even if it's over the pass.
I've heard the Enchantments are nice.
I don't want the hike to be over 10 miles round trip. I'd consider a point-to-point hike.
I want the hike to be as seculuded as possible. I don't want to see many people.
I would like it to be as bug free as possible. I would also like to avoid any Forest Service roads.
""";

Run the app with dotnet run and see what it recommends now. Here's an example we received from one run:

Great! Based on your preferences, I'd recommend the following strenuous hikes with amazing views near the I-90 corridor:

1. Mt. Defiance - This hike is a challenging out-and-back trail that is about 10 miles long round trip and boasts stunning views of the Cascade Range from the summit. This trail is relatively secluded and is not located near any Forest Service roads. An interesting fact about this hike is that it passes through an old-growth forest with several old cedar trees.

2. Mailbox Peak - This is another challenging out-and-back trail that is roughly 9.4 miles long round trip. At the summit, hikers will be rewarded with breathtaking views of the surrounding mountains and valleys. Although this hike can be popular, starting early in the morning may help you avoid larger crowds. Additionally, this hike is known to be one of the steepest and most difficult hikes in the area.

3. Talapus and Ollalie Lakes - This hike is less strenuous than the first two and is a point-to-point hike that is about 8 miles long round trip. Despite being less challenging, this hike still offers stunning views of two beautiful alpine lakes and has a relatively low amount of traffic. Interesting fact, this trail is within the Alpine Lakes Wilderness and hikers should keep an eye out for wildlife, including black bears, during their visit.

Hope this helps and happy hiking!

This is much better! The response is following the rules of the system prompt and providing detailed information of the hike description requested.

If you look closely, hike recommendation number 3 says it is both a point-to-point and a round trip hike. Let's change recommendation request prompt to see if we can avoid that mis-information.

Chain-of-thought

Asking the model to provide a step-by-step accounting of its thought process should help it recommend even more accurate recommendations.

Update the hikeRequest variable to be the following:

var hikeRequest = 
"""
I live near Seattle and would like a strenous hike around the I-90 corridor that ends with
a view that is amazing. I don't mind driving for a long time, even if it's over the pass.
I've heard the Enchantments are nice.
I don't want the hike to be over 10 miles round trip. I'd consider a point-to-point hike.
I want the hike to be as seculuded as possible. I don't want to see many people.
I would like it to be as bug free as possible. I would also like to avoid any Forest Service roads.

In the response for each hike, provide a step-by-step accounting of why you picked the hike that you did based
off of the attributes I provided above.
""";