top of page
Writer's pictureKrishna Bhatt

The Art of Prompt Engineering: Guiding AI to Better Outputs

In the bustling realm of artificial intelligence, where machines are learning to write, reason, and even 'think' in ways mirroring human cognition, the manner in which we communicate with these algorithms plays a pivotal role. Enter the fascinating world of "Prompt Engineering"—a discipline that, although less highlighted, serves as the guiding star in our interactions with powerful language models like OpenAI's GPT series.


At its core, prompt engineering is about crafting the perfect question or statement to extract the desired response from an AI, ensuring precision, relevance, and clarity. Think of it as the GPS for navigating the vast intelligence of AI models.


This blog delves deep into the nuances of prompt engineering, exploring its significance, techniques, and best practices. Whether you're an AI enthusiast, a developer, or simply curious, join us as we demystify the art of asking the right questions in the AI era.

What is prompt engineering

Prompt engineering is akin to providing clear and precise instructions to sophisticated computer programs, known as LLMs, to optimize their performance. Just as providing clearer directions ensures a friend doesn’t lose their way while driving, an effective prompt can guide an LLM to produce more accurate and relevant results. LLMs are powerful computer models capable of generating text, images, and other content, but their outputs largely depend on the clarity and specificity of the prompts they receive. For instance:

  1. If you need a description of a product, giving the LLM detailed features and benefits about the product can result in a more precise and compelling description.

  2. Thoughtfully crafted prompts can even help reduce bias and make the output more fair and inclusive.

However, it's crucial to remember that, regardless of how refined a prompt is, outputs from these AI models aren't infallible. They might still carry biases or inaccuracies, hence the emphasis on always using AI responsibly and verifying its outputs.


One of the major advantages of prompt engineering is that it offers insights into the model's thought process. Especially in crucial sectors like healthcare, understanding how decisions are made is of paramount importance. To make prompts even better, there are various techniques such as:

  • Providing instructions

  • Introducing contextual content

  • Giving cues or "few-shot examples"

  • Correctly ordering content

Prompt engineering is a dynamic field, continually evolving with new methods being explored, and it plays a pivotal role in harnessing the full potential of LLMs.


Azure OpenAI and OpenAI Endpoint Considerations

When delving into how prompt engineering can enhance the performance of models within Azure OpenAI and OpenAI, it's pivotal to understand the distinctions between various API endpoints and their respective uses. Both the Completion and ChatCompletion endpoints can yield similar outcomes, but ChatCompletion stands out due to its greater adaptability in crafting prompts, especially tailored for chat contexts.

Functionally, the ChatCompletion endpoint offers a unique feature: the ability to define a system message for the AI model. Furthermore, it's structured to incorporate prior messages within the prompt. Conversely, when working with the Completion endpoint, this kind of utility can be emulated using a mechanism known as a meta prompt. Comprehensive insights into these facets are provided in subsequent units.

Model-wise, both endpoints support a range of models, such as gpt-35-turbo. However, the gpt-4 generation models are exclusively compatible with ChatCompletion.

While the Completion endpoint is capable of producing comparable outcomes, it necessitates meticulous prompt formatting to ensure optimal comprehension by the AI model. Most examples in this module are apt for ChatCompletion, but they can be adapted for Completion.

An additional point to remember is that ChatCompletion isn't restricted to chat-based applications. It can also be employed in other scenarios where directives are housed in the system message and user input is encapsulated in the user role message.


Using the Completion Endpoint

Adjusting model Settings


Changing some model settings can change how it answers. The settings called temperature and top_p are important. They control how "random" or "creative" the model's answers can be.

If you set these values high, the model gives creative answers. They might be fun but not always on point. If you set them low, the model gives straight and clear answers.

Test and change these settings one at a time with the same question to see how the answers change. It's a good idea to change only one setting at once, not both together.


Here are all the paramers of OpenAI



Write more effective prompts


OpenAI's tools are great at answering questions. But to get the best answers, we need to ask the right way. Just like in a game of "20 Questions", if we ask better, we get better answers. So, if developers learn how to ask questions the right way to OpenAI, they'll get clearer and more helpful answers.


Provide clear instructions


When you're clear with what you ask the OpenAI tool, you're more likely to get the answer you want. The clearer and more detailed you are, the better the tool can answer your question. Let's say you're trying to describe a new water bottle. Think about how the answer might change based on how you ask:


Try same prompt with clear instructions :

When you tell the model exactly what you want, like giving it a list of details or saying how long you want the answer to be, it helps the model create a better description for the new product. So, just tell the model what you need, and you'll get a more accurate result.


Format of instructions


How you structure your instructions is important because it can change how the model understands what you want. Sometimes, the model pays more attention to the stuff at the end of your request than the beginning. To make sure you get the best response, try repeating your instructions at the end and see if that helps.


This also matters when you're having a chat with the model. What you or the model said recently in the conversation can affect how it responds. In the next part of this guide, we'll dig deeper into using conversations to get better answers. But for now, remember that putting important information towards the end of your request can lead to a better response.


Use section markers


You can make your instructions clearer by putting them at the beginning or end of your message, and separate them from the rest of your message using lines like this: "---" or "###". This helps the model understand what's an instruction and what's the actual content you want.


Primary, supporting, and grounding content


When you want the model to give a good answer, you need to provide it with some information. There are two types of information you can give: primary and supporting content.


Primary content is the main stuff you want the model to focus on. It could be a sentence you want it to translate or a whole article you want it to summarize. You usually put this at the beginning or end of your message, inside special blocks like "---", and you tell the model what you want it to do with that content.

For example, if you have a long article you want the model to summarize, you can put it in a "---" block in your message and then tell the model to summarize it.


Supporting content is content that may alter the response, but isn't the focus or subject of the prompt. Examples of supporting content include things like names, preferences, future date to include in the response, and so on. Providing supporting content allows the model to respond more completely, accurately, and be more likely to include the desired information.


For example, given a very long promotional email, the model is able to extract key information. If you then add supporting content to the prompt specifying something specific you're looking for, the model can provide a more useful response. In this case the email is the primary content, with the specifics of what you're interested in as the supporting content



"Grounding content" is like giving the model a good source of information to help it give better answers. This source could be things like an essay, an article, or a company's frequently asked questions (FAQs). It can also be information that's more up-to-date than what the model already knows.


When you want the model to give reliable and current answers, or you need it to use information that's not widely known, you should use grounding content.


This is different from "primary content," which is what the model works on directly, like when you ask it to summarize a paragraph. Grounding content is like the library of information the model uses to answer your questions. For example, if you give the model a research paper on AI history that hasn't been published yet, it can use that paper to answer your questions.


Prompt


Cues

In the context of prompt engineering, "cues" refer to specific words, phrases, or formatting techniques that you use within your prompt to guide the AI model in generating the desired response. Cues are essentially signals that tell the model how to approach the task or question you've presented.


For example, if you want to use a cue to instruct the model to provide a detailed answer, you might include phrases like "Please explain in detail..." or "Can you elaborate on...". These cues signal to the model that you're looking for a comprehensive response.


Cues can be particularly important in fine-tuning the AI's output, as they help clarify your intentions and steer the model toward the specific information or style you want in the response. They can also help mitigate issues like biases or inaccuracies in the AI's responses by providing clear context and guidance.


In the context of SQL queries, cues are specific instructions or formatting techniques used in your input text to guide the AI model in generating the desired SQL query or database-related response. These cues help the model understand the structure and purpose of your query.


The model response picks up where the prompt left off, continuing in SQL, even though we never asked for a specific language.


Provide context to improve accuracy

By providing context to the AI model, it allows the model to better understand what you are asking for or what it should know to provide the best answer. Context can be provided in several ways.


Request output composition

Specifying the structure of your output can have a large impact on your results. This could include something like asking the model to cite their sources, write the response as an email, format the response as a SQL query, classify sentiment into a specific structure, and so on. For example:


Prompt:

Write a table in markdown with 6 fruits in it, with their genus and species.



This technique can be used with custom formats, such as a JSON structure:


Prompt:

Put 6 fruits with their genus and species in JSON


System message

In prompt engineering, a "System message" refers to a message or instruction provided at the beginning of a conversation or prompt to set the context or behavior of the AI model. It is typically used to guide the AI's behavior in a chat or dialogue scenario.

Here's an explanation of the key aspects of a System message:

  1. Context Setting: A System message is used to establish the context of the conversation or prompt. It can provide information about the scenario, the role of the AI, or any specific instructions that should apply throughout the conversation.

  2. Role Definition: The System message can define the role of the AI model. For example, it can specify that the AI is acting as a tutor, a virtual assistant, or a creative writer, depending on the desired interaction.

  3. Conversation Flow: System messages can influence the flow of the conversation by instructing the AI model on how to respond to user inputs. For instance, it can instruct the AI to answer questions, provide explanations, or engage in storytelling.

  4. Instructions: System messages often include explicit instructions for the AI, such as asking it to think step by step, brainstorm ideas, or provide detailed explanations.

  5. Formatting: In chat-based interactions, System messages are usually formatted differently from user messages, making it clear to the AI model that this message sets the stage for the conversation.

  6. Styling and Persona: System messages can also be used to define the tone, style, or persona of the AI, whether it should be formal, friendly, or creative.

Here's an example of a System message in a conversation:


Prompt:

System: Welcome to the TravelBot! I'm here to assist you with travel-related questions and recommendations. Feel free to ask about flights, hotels, or places to visit. If you need information about a specific destination, just let me know!


In this example, the System message sets the context by introducing the AI as a travel assistant and provides guidance on the types of questions the user can ask.


Overall, System messages are a crucial component of prompt engineering in chat-based AI interactions, as they establish the groundwork for how the AI model should engage with the user and ensure a coherent and useful conversation.


Chat History

When you talk to AI, you can include previous messages in the conversation to help the AI understand and respond better. This is called "chat history."

There are two ways to provide this chat history:

  1. You can show the AI a real chat conversation you've had before.

  2. You can make up an example conversation.

When you use chat systems like ChatGPT, they automatically keep track of what's been said before. This makes the conversation more useful and interesting.

In some chat systems, like the one in Azure OpenAI Studio, you can decide how much of the chat history you want the AI to see. You can choose to show just the most recent message or a lot of messages from the past.


Keep in mind that the more chat history you include, the more "tokens" you use. Tokens are like units of text, and there's a limit to how many a model can handle. So, you need to balance how much chat history you include based on the model's limits.


Some chat systems can also summarize the chat history to save on tokens. This means they make a short version of what was said before and only show a few recent messages exactly as they were said


Few shot learning

Few-shot learning in AI prompt engineering is like teaching the AI with a small number of examples to make it better at responding to questions or requests. For instance, imagine you're training an AI to recognize sentiments (positive or negative) in sentences. You give it a few examples like this:

  • User: That was an awesome experience Assistant: positive

  • User: I won't do that again Assistant: negative

  • User: That was not worth my time Assistant: negative

NODEJS


Here, the AI learns that when it sees positive words, it should respond with "positive," and when it sees negative words, it should respond with "negative." It's like training a pet to react to certain words or actions.

But, if you only give the AI a sentence like "You can't miss this" without any more context, it might not know what to do because it didn't see that specific example before.

So, in few-shot learning, you teach the AI by showing it a few examples, and it learns from those to respond better. It's like giving it a little taste of different situations so it can handle similar ones in the future.

Break down a complex task


Breaking down a complex task means taking a big problem and splitting it into smaller, easier-to-handle parts. This helps the AI understand each piece better and makes its answers more accurate. It also lets you use what the AI said before in the conversation to get even better answers.

For example, let's say you have a question like, "How far can Sally drive in 5 hours?" Instead of asking it all at once, you can break it down. First, you might ask, "How fast does Sally drive?" Then, after you get the answer, you can ask, "If Sally drives at that speed for 5 hours, how far will she go?"

By breaking it into two questions, you get a more precise answer because you considered the speed first. It's like solving a big puzzle one piece at a time to see the whole picture.

Chain of thought


The "chain of thought" method involves asking an AI model to explain how it thinks step by step. It's like asking the AI to show its work, just as you would in a math class. Instead of simply getting an answer, you can request the AI to break down its reasoning process, which can be incredibly useful.

Let's say you ask an AI model, "Which sport is the easiest to learn but the hardest to master?" Rather than just receiving a single answer, you can instruct the AI to walk you through its thought process step by step.

In response, the AI might explain:

  1. "To begin, I considered a list of popular sports."

  2. "Next, I evaluated how straightforward it is to grasp the fundamentals of each sport."

  3. "I also weighed the level of difficulty associated with achieving mastery in each sport."

  4. "Then, I compared the ease of initial learning to the challenge of reaching a high level of proficiency."

  5. "Based on this analysis, I arrived at the conclusion that golf is the sport that fits the criteria of being the easiest to learn yet the hardest to master."

This approach allows you to gain insight into how the AI arrived at its answer. If the AI makes a mistake or reaches an incorrect conclusion, you can pinpoint where it went wrong and make necessary adjustments to your query.


For example, if the AI mistakenly suggests chess as the easiest sport to learn but the hardest to master, you can recognize the error and provide clarification or ask a more specific question to obtain a more accurate response.


In essence, the "chain of thought" method empowers you to refine your questions, enhance the AI's understanding, and ultimately obtain more precise and insightful answers. It's akin to reviewing an AI model's thought process to ensure its responses are on the right track.

Take Away

In conclusion, AI prompt engineering is the art of crafting precise and effective instructions to guide AI models in generating desired responses. It's not merely a technical task but a creative endeavor that requires the human touch. To harness the full potential of AI models, one must not only understand the technology but also be a creative and critical thinker.


AI prompt engineering is a dynamic field where innovation and ingenuity play a pivotal role. Crafting prompts that are clear, context-rich, and tailored to the specific task can lead to remarkable outcomes. The ability to anticipate how an AI model interprets instructions and creatively construct prompts can yield responses that are insightful, accurate, and contextually relevant.


Being a masterful prompt engineer means knowing the technology inside and out, but it also means thinking outside the box. It involves exploring new approaches, experimenting with different phrasings, and continuously refining prompts to achieve optimal results. By marrying technical expertise with creativity and critical thinking, we can unlock the true potential of AI models and push the boundaries of what they can accomplish.


In this ever-evolving landscape of artificial intelligence, prompt engineering is both a science and an art. It's a journey of discovery and innovation where human intelligence and machine intelligence converge to create something extraordinary. As we continue to advance in this field, let's remember that the most successful prompt engineers are those who not only know what AI technology is but also dare to imagine what it can be.




53 views2 comments

Recent Posts

See All

2 Comments

Rated 0 out of 5 stars.
No ratings yet

Add a rating
Guest
Sep 16, 2023
Rated 5 out of 5 stars.

Excellently written article on the art of Prompt Engineering!

Like

Guest
Sep 16, 2023
Rated 5 out of 5 stars.

well written

Like
bottom of page