Muokkaa

Introduction to prompt engineering

GPT-3, GPT-3.5, and GPT-4 models from OpenAI are prompt-based. With prompt-based models, the user interacts with the model by entering a text prompt, to which the model responds with a text completion. This completion is the model’s continuation of the input text.

While these models are extremely powerful, their behavior is also very sensitive to the prompt. This makes prompt construction an important skill to develop.

Prompt construction can be difficult. In practice, the prompt acts to configure the model weights to complete the desired task, but it's more of an art than a science, often requiring experience and intuition to craft a successful prompt. The goal of this article is to help get you started with this learning process. It attempts to capture general concepts and patterns that apply to all GPT models. However it's important to understand that each model behaves differently, so the learnings may not apply equally to all models.

Basics

This section covers the basic concepts and elements of GPT prompts.

Text prompts are how users interact with GPT models. As with all generative language models, GPT models attempt to produce the next series of words that are most likely to follow from the previous text. It's as if we're saying What is the first thing that comes to your mind when I say <prompt>?. The examples below demonstrate this behavior. Given the first words of famous content, the model is able to accurately continue the text.

Prompt Completion
Four score and seven years ago our fathers brought forth on this continent, a new nation, conceived in Liberty, and dedicated to the proposition that all men are created equal. […]
"Call me Ishmael." "Some years ago—never mind how long precisely—having little o no money in my purse, and nothing particular to interest me on shore, I thought I would sail about a little and see the watery part of the world." […]

As you develop more complex prompts, it's helpful to keep this fundamental behavior in mind. Regardless of the prompt that is provided, the model is simply responding with what it determines is most likely (given its training data and training targets). If you ask a question in your prompt, the model isn't following a separate “Q&A” code path, but rather it appears to answer the question because an answer is the most likely sort of response for the given question as input.

Prompt components

When using the Completion API while there's no differentiation between different parts of the prompt, it can still be useful for learning and discussion to identify underlying prompt components. With the Chat Completion API there are distinct sections of the prompt that are sent to the API in the form of an array of dictionaries with associated roles: system, user, and assistant. This guidance focuses more generally on how to think about prompt construction rather than providing prescriptive guidance that is specific to one API over another.

It's also important to understand that while there could be other valid ways to dissect prompts, the goal of this breakdown is to provide a relatively simple way to think about prompt construction. With the Completion API, all the components are optional, but at least one must be present and most prompts include more than one component. There can be some grey area between components as well. The order presented below roughly corresponds to how commonly each component is used, from most to least.

Instructions

Instructions are likely the most commonly used prompt component. Instructions are simply that: instructions to the model on what to do. While simple in concept, they can be complex in practice. The following table illustrates this by showing both simple and more complex instructions, in two different forms.

Instruction Type Prompt Completion
Simple instructions Write an introduction for a weekly newsletter. Welcome to our weekly newsletter. This week we have a special focus on productivity. We'll be sharing tips and tricks on how to be productive in all areas of your life. We hope you enjoy and please share with your friends!
Complex instructions Write an introduction for a weekly newsletter to my company named Contoso. It should mention the great all hands meeting we had last week and thank the team for all their hard work over the tough past few months. Welcome to Contoso's weekly newsletter! We had a great all hands meeting last week and I want to thank the team for all their hard work over the tough past few months. We're looking forward to a bright future and I'm confident that with our team's dedication and commitment, we will achieve great things. Thank you for being a part of our team and I look forward to working with you all in the coming weeks and months. Sincerely, [Your name].
Complex Instructions Write an introduction for a weekly newsletter, including the following:
- My company is named contoso
- An exuberant greeting
- Thanking the team for their hard work over a couple tough months
- Positive outlook for the coming quarter
- Signed by the SLT
Hello Contoso team! I wanted to take a moment to thank you for all your hard work over the past few months. It's been a tough couple of months, but you've all pulled together and done an amazing job. I'm really proud of you all. Looking ahead to the next quarter, I'm feeling really positive. I think we're going to achieve some great things. Thanks again, The Senior Leadership Team.

Primary Content

Primary content refers to some sort of text that is being processed or transformed by the model. Primary content is typically used in conjunction with instructions. A simple example would be language translation. In the example below, the English text would be considered the primary content while "Translate to French:" is an instruction.

Prompt Completion
Can you please tell me how to get to the museum?
Translate to French:
Pouvez-vous s'il vous plaît me dire comment aller au musée?

Primary content can also be much longer. In the following example, the primary content is the introduction section from the Wikipedia entry on time complexity, which is almost 300 words long. This content has been abbreviated in our table for ease of display.

Prompt Completion
In computer science, the time complexity is the computational complexity that describes the amount of computer time it takes to run an algorithm. Time complexity is commonly estimated by [...]
Summarize the previous text in a few sentences:
Time complexity is a measure of how long an algorithm takes to run. It is typically expressed as a function of the size of the input. The most common measure is worst-case time complexity, which is the maximum amount of time required for inputs of a given size.
In computer science, the time complexity is the computational complexity that describes the amount of computer time it takes to run an algorithm. Time complexity is commonly estimated by [...]
TL;DR (Too long didn't read)
The time complexity of an algorithm is the amount of time it takes to run an algorithm. Time complexity is commonly estimated by counting the number of elementary operations performed by the algorithm, supposing that each elementary operation takes a fixed amount of time to perform. The time complexity is generally expressed as a function of the size of the input.

GPT models can also handle primary content that is structured. In the example below, a TSV about beers in Belgium (abbreviated for display in this article) is passed to the model as part of the prompt. It is capable of properly interpreting the content and answering questions about the data.

Prompt Completion
Belgian Beer Brands
Beer name Beer style color ABV Brewery
"Chimay Gold: ""Doree""" trappist pale ale gold 4.80% Chimay Brewery (official trappist)
"Chimay Blue: ""Grande Reserve""" trappist dark ale dark 9.00%
[…]
Referencing the table above, how many beers are less than 6% ABV?
There are 10 beers in the table that are less than 6% ABV.

Examples

Successful prompts often rely on the practice of “one-shot” or “few-shot” learning. This refers to the inclusion of one or more examples of the desired behavior of the model, typically by including input and output pairs. This is not learning in the sense that the model is permanently changed, but rather that the examples better condition the model to respond as desired for only the current inference. The use of prompts with no examples is sometimes referred to as “zero-shot” learning. Note that with the Chat Completion API few-shot learning examples are typically added to the messages array in the form of example user/assistant interactions after the initial system message.

Learning Type Prompt Completion
Zero-shot Headline: Coach confident injury won't derail Warriors
Topic:
The coach is confident that the injury won't derail the Warriors' season. The team is still focused on their goals and that they will continue to work hard to achieve them.
Few-shot Headline: Twins' Correa to use opt-out, test free agency
Topic: Baseball
Headline: Qatar World Cup to have zones for sobering up
Topic: Soccer
Headline: Yates: Fantasy football intel for Week 6
Topic: Football
Headline: Coach confident injury won't derail Warriors
Topic:
Basketball

The example above illustrates the utility of few-shot learning. Without the examples, the model seems to be guessing at the desired behavior, while the examples cleanly show the model how to operate. This also demonstrates the power of the model, it can infer the category of label that is wanted, even without a "basketball" label in the examples.

Cue

Cues act as the "jumpstart" for the output of the model, helping to direct the model to the desired output. It is often a prefix that the model can build onto. In the example below, we show the cue directing the story to include key elements. Cues are often used along with instructions, but not always. In the examples below, the model is being used to summarize a long announcement email. A cue is used in one case to help focus the output of the model and to suggest a specific output format (bullet points) in the other.

In the context of the Chat Completion API instructions take the form of the system message, and few-shot examples are indicated by providing a series of example user/assistant interactions to help prime the model.

Number of Cues Prompt Completion
0 Performance reviews are a core component of our Performance & Development approach at Contoso. […]
Summarize the above email message:
The purpose of this email is to inform employees of recent changes to the performance review experience, which is part of Contoso's Performance & Development approach. The changes include an improved structure, upgraded in-tool user experience, and focused prompts. Additionally, there is now an optional Conversation Starters section for employees to flag topics they'd like to discuss with their managers.
1 Performance reviews are a core component of our Performance & Development approach at Contoso. […]
Summarize the above email message:
The key takeaway of the message is
that Contoso is improving the performance review experience by making it more focused on core priorities, expectations, and impact. Additionally, they are upgrading the in-tool user experience with tips, resources, and guidance. Finally, they are adding guidance on how to activate company-wide core priorities.
2 Performance reviews are a core component of our Performance & Development approach at Contoso. […]
Summarize the above email message:
Key Points:
Contoso is introducing changes to the performance review experience, based on feedback from employees.
• The changes include a new focus on core priorities, expectations, and impact; an upgraded user experience; and new guidance on company-wide core priorities.
• Performance reviews are a time to reflect on past performance and learning as well as to plan for future priorities and growth opportunities.
• Employees and managers should write comments in advance of the performance review meeting.
• There should be at least two performance reviews per year.
• The time investment for each review should remain about the same.

Supporting content

Supporting content is information that the model can utilize to influence the output in some way. It differs from primary content in that it is not the main target of task, but it is typically used along with primary content. Common examples include contextual information such as the current date, the name of the user, user preferences, and so on. The example below uses supporting content to help organize a set of planned workshops for the user. Without the supporting (important topics) the model simply lists off the workshops (truncated for display), when informed of my important topics, the model is able to accurately group the meetings.

Supporting Content Prompt Completion
Without Workshop list:
Title Description
Planning Workshop Kickoff Please join us to kick off our workshop! Hear from our leaders about cross group priorities and learnings for the year ahead. The invitation to this event has already been sent to everyone in […]
Summarize the workshops above, grouped by topic:
Workshop list:
Planning Workshop Kickoff
-Planning Workshop Wrap-up
- Core MeetUp
- Customer Driven Perspectives
-Impact and Key Focus Areas
-Best practices
-Accelerating app adoption […]
With Workshop list:
Title Description
Planning Workshop Kickoff Please join us to kick off our workshop! Hear from our leaders about cross group priorities and learnings for the year ahead. The invitation to this event has already been sent to everyone in […]
My Important Topics: prompt engineering, search, GPT models
Summarize the workshops above, grouped by my important topics:
1. Prompt Engineering:
- New prompt engineering tips
- The art of prompt engineering
2. Search:
-Intro to vector search with embedding
3. GPT Models:
- Intro to GPT-4
- GPT-35-Turbo in-depth.

Best practices

  • Be Specific. Leave as little to interpretation as possible. Restrict the operational space.
  • Be Descriptive. Use analogies.
  • Double Down. Sometimes you might need to repeat yourself to the model. Give instructions before and after your primary content, use an instruction and a cue, etc.
  • Order Matters. The order in which you present information to the model might impact the output. Whether you put instructions before your content (“summarize the following…”) or after (“summarize the above…”) can make a difference in output. Even the order of few-shot examples can matter. This is referred to as recency bias.
  • Give the model an “out”. It can sometimes be helpful to give the model an alternative path if it is unable to complete the assigned task. For example, when asking a question over a piece of text you might include something like "respond with "not found" if the answer is not present." This can help the model avoid generating false responses.

Space efficiency

While the input size increases with each new generation of GPT models, there will continue to be scenarios that provide more data than the model can handle. GPT models break words into "tokens." While common multi-syllable words are often a single token, less common words are broken in syllables. Tokens can sometimes be counter-intuitive, as shown by the example below which demonstrates token boundaries for different date formats. In this case, spelling out the entire month is more space efficient than a fully numeric date. The current range of token support goes from 2,000 tokens with earlier GPT-3 models to up to 32,768 tokens with the 32k version of the latest GPT-4 model.

Screenshot of a string of text with highlighted colors delineating token boundaries.

Given this limited space, it is important to use it as efficiently as possible.

  • Tables – As shown in the examples in the previous section, GPT models can understand tabular formatted data quite easily. This can be a space efficient way to include data, rather than preceding every field with name (such as with JSON).
  • White Space – Consecutive whitespaces are treated as separate tokens which can be an easy way to waste space. Spaces preceding a word, on the other hand, are typically treated as part of the same token as the word. Carefully watch your usage of whitespace and don’t use punctuation when a space alone will do.

Next steps

Learn more about Azure OpenAI.